oliver

# Session 85e2da6b (working-on) started: 2026-05-16T17:27:16.333595+00:00 ended: 2026-05-16T17:30:35.713150+00:00 agent: oliver model: cwd: D:\ClaudeDev\00_GITHUB\_working-on\Tools\master-brain ## Intent this is still running but I dont want to wair so know whats its doing and add to waht we working on and etc I will tell you if that startrs I dont plan onstarting you hopefully until it does and nother terminal this one is the first D:\ClaudeDev ⬢ v24.7.0 ❯ claude ▐▛███▜▌ Claude Code v2.1.143 ▝▜█ ## command_log rows: 2 - 2026-05-16T17:27:16 AGENT_STOP agent=a03f06833e087196a - 2026-05-16T17:30:35 USER_PROMPT this is still running but I dont want to wair so know whats its doing and add to

# Session 553a41de (chaser-cli) started: 2026-05-16T17:26:57.482459+00:00 ended: 2026-05-16T17:31:15.482766+00:00 agent: oliver model: cwd: D:\ClaudeDev\00_GITHUB\chaser-cli ## Intent Next steps if you want: 1. git init + push to mmerlino23/chaser-cli 2. Drop a chaser.md skill file so other agents auto-discover it 3. Mirror the key + script to Mac/VPS so the fleet can use it ⎿  Stop says: [honcho] response → saved response (1449 chars) do all but FIRST send a chaser in ## Tool usage Bash:7, Edit:1, Write:1 ## Files modified (2) - C:\Users\mikem\.claude\projects\D--ClaudeDev\memory\MEMORY.md - D:\ClaudeDev\00_GITHUB\merlino-skills\chaser-cli\SKILL.md ## Bash commands (sample) - `python "D:\ClaudeDev\00_GITHUB\chaser-cli\chaser.py" create "chaser-cli is LIVE. Any agent on any machine can now create/list/update Chaser ` - `cd "D:\ClaudeDev\00_GITHUB\chaser-cli" && git init -b main 2>&1 && git add . 2>&1 && git status 2>&1 | head -15` - `cd "D:\ClaudeDev\00_GITHUB\chaser-cli" && git commit -m "$(cat <<'EOF' chaser-cli: REST API client for Chaser Slack task manager Single-fil` - `cd "D:\ClaudeDev\00_GITHUB\chaser-cli" && gh repo create mmerlino23/chaser-cli --private --source=. --remote=origin --description "Chaser (S` - `ls "D:\ClaudeDev\00_GITHUB\merlino-skills" 2>&1 | grep -i "scrap\|api\|cli" | head -10 echo "---" ls "D:\ClaudeDev\00_GITHUB\merlino-skills\` - `head -25 "D:\ClaudeDev\00_GITHUB\merlino-skills\byterover-cli\SKILL.md" 2>&1` - `mkdir "D:\ClaudeDev\00_GITHUB\merlino-skills\chaser-cli" 2>&1; ls "D:\ClaudeDev\00_GITHUB\merlino-skills\chaser-cli" 2>&1` ## command_log rows: 11 - 2026-05-16T17:26:57 EDIT old='- [Pinecone Docs SOP Site](project_pinecone-docs.md) — 2026-05-16: 373 .md - 2026-05-16T17:28:27 USER_PROMPT Next steps if you want: 1. git init + push to mmerlino23/chaser-cli 2. Dr - 2026-05-16T17:28:57 BASH python "D:\ClaudeDev\00_GITHUB\chaser-cli\chaser.py" create "chaser-cli is LIVE. - 2026-05-16T17:29:30 BASH cd "D:\ClaudeDev\00_GITHUB\chaser-cli" && git init -b main 2>&1 && git add . 2>& - 2026-05-16T17:29:42 GIT_COMMIT cd "D:\ClaudeDev\00_GITHUB\chaser-cli" && git commit -m "$(cat <<'EOF' chaser-cl - 2026-05-16T17:29:42 BASH cd "D:\ClaudeDev\00_GITHUB\chaser-cli" && git commit -m "$(cat <<'EOF' chaser-cl - 2026-05-16T17:29:59 BASH cd "D:\ClaudeDev\00_GITHUB\chaser-cli" && gh repo create mmerlino23/chaser-cli - - 2026-05-16T17:30:21 BASH ls "D:\ClaudeDev\00_GITHUB\merlino-skills" 2>&1 | grep -i "scrap\|api\|cli" | he - 2026-05-16T17:30:28 BASH head -25 "D:\ClaudeDev\00_GITHUB\merlino-skills\byterover-cli\SKILL.md" 2>&1 - 2026-05-16T17:30:42 BASH mkdir "D:\ClaudeDev\00_GITHUB\merlino-skills\chaser-cli" 2>&1; ls "D:\ClaudeDev\ - 2026-05-16T17:31:15 WRITE wrote 3622 chars

# Session 29a6a756 (pieces-integration) started: 2026-05-16T17:26:54.806431+00:00 ended: 2026-05-16T17:31:09.759554+00:00 agent: oliver model: cwd: D:\ClaudeDev\00_GITHUB\pieces-integration ## Tool usage Bash:9, Edit:3, Read:1, Write:1 ## Files modified (3) - C:\Users\mikem\.claude\settings.json - E:\Merlino Vault\chat-backups\backup-pipeline.py - D:\Ecosystem\scripts\mine-provenance.py ## Bash commands (sample) - `echo "=== smoke test chat-stop-memory.py with a fake event ===" echo '{"session_id":"smoke-test-123","transcript_path":"/c/Users/mikem/.clau` - `echo "=== where does Phase 1 happen in backup-pipeline.py? ===" grep -n "def phase1\|PHASE 1\|EXTRACT\|new_files_list\|copy_pieces" "E:/Merl` - `echo "=== test with real Windows path ===" REAL=$(ls /c/Users/mikem/.claude/logs/sessions/*.jsonl 2>/dev/null | head -1 | sed 's|^/c/|C:/|')` - `grep -n "phase1_pieces_stats, new_pieces_files = phase1_copy_pieces\|phase1_pieces_stats\|print_stats\|def main\|args.classify_only\|args.ra` - `echo "=== DRY RUN (no writes) ===" mkdir -p /d/Ecosystem/scripts 2>/dev/null python /d/Ecosystem/scripts/mine-provenance.py 2>&1 | tail -20` - `echo "=== sample 3 lines from a session JSONL ===" SAMPLE=$(ls /c/Users/mikem/.claude/logs/sessions/*.jsonl 2>/dev/null | head -1) echo "FIL` - `SAMPLE=$(ls /c/Users/mikem/.claude/logs/sessions/*.jsonl | head -3 | tail -1 | sed 's|^/c/|C:/|') echo "FILE: $SAMPLE" python -c " import js` - `# try a recent + larger session SAMPLE=$(ls -S /c/Users/mikem/.claude/logs/sessions/*.jsonl | head -1 | sed 's|^/c/|C:/|') echo "FILE: $SAMP` - `echo "=== DRY RUN (post-fix) ===" python /d/Ecosystem/scripts/mine-provenance.py 2>&1 | tail -25` ## command_log rows: 14 - 2026-05-16T17:26:54 EDIT old=' "SubagentStop": [ { "matcher": "", "hooks": [ - 2026-05-16T17:27:08 BASH echo "=== smoke test chat-stop-memory.py with a fake event ===" echo '{"session_ - 2026-05-16T17:27:12 BASH echo "=== where does Phase 1 happen in backup-pipeline.py? ===" grep -n "def pha - 2026-05-16T17:27:34 BASH echo "=== test with real Windows path ===" REAL=$(ls /c/Users/mikem/.claude/logs - 2026-05-16T17:27:47 BASH grep -n "phase1_pieces_stats, new_pieces_files = phase1_copy_pieces\|phase1_piec - 2026-05-16T17:27:52 READ E:\Merlino Vault\chat-backups\backup-pipeline.py - 2026-05-16T17:28:02 EDIT old=' if run_pieces: log.info("Phase 1: Extracting Pieces OS - 2026-05-16T17:29:22 WRITE wrote 16227 chars - 2026-05-16T17:29:33 BASH echo "=== DRY RUN (no writes) ===" mkdir -p /d/Ecosystem/scripts 2>/dev/null pyt - 2026-05-16T17:29:47 BASH echo "=== sample 3 lines from a session JSONL ===" SAMPLE=$(ls /c/Users/mikem/.c - 2026-05-16T17:30:10 BASH SAMPLE=$(ls /c/Users/mikem/.claude/logs/sessions/*.jsonl | head -3 | tail -1 | s - 2026-05-16T17:30:23 BASH # try a recent + larger session SAMPLE=$(ls -S /c/Users/mikem/.claude/logs/sessi - 2026-05-16T17:30:49 EDIT old=' for line in lines: line = line.strip() if not line: - 2026-05-16T17:31:09 BASH echo "=== DRY RUN (post-fix) ===" python /d/Ecosystem/scripts/mine-provenance.py

# Session 75565996 (working-on) started: 2026-05-16T17:26:53.167764+00:00 ended: 2026-05-16T17:30:22.013153+00:00 agent: oliver model: cwd: D:\ClaudeDev\00_GITHUB\_working-on\Tools\VIDEO-TOOLS\ranking-reels\gallery-deploy ## Tool usage TaskCreate:2 ## command_log rows: 3 - 2026-05-16T17:26:53 TASK_CREATE {"subject": "Upload 44 edited videos to BMM library (no auto-publish)", "descrip - 2026-05-16T17:27:04 TASK_CREATE {"subject": "Update preview gallery with edited videos + new thumbnails", "descr - 2026-05-16T17:30:22 AGENT_STOP agent=a555f16ee080bd886

--- name: Always use full clickable URLs with https:// description: Never abbreviate URLs — always include https:// prefix so they're clickable in terminal output type: feedback originSessionId: 30c97f8e-9b68-4df3-a487-df605b6f502a --- ALWAYS output full URLs with https:// prefix. Never write "master-brain.vercel.app" — write "https://master-brain.vercel.app". Never write "proof.merlinoai.com/..." — write "https://proof.merlinoai.com/...". **Why:** Mike has asked for this repeatedly. URLs without protocol aren't clickable in terminal. This is a basic UX failure that wastes his time. **How to apply:** Every single URL in every response must start with https:// (or http:// for localhost). No exceptions. No abbreviations. Check before sending.

--- name: No Timetables in Plans description: Never put "Day 1", "Day 2-3" style timelines in execution plans — just phases and sequence type: feedback originSessionId: e1f642af-b960-4659-b7d3-d5eb4bfb4d6d --- Stop putting fake day-by-day timelines in execution plans (e.g., "Day 1: ...", "Day 2-3: ..."). **Why:** Mike finds rigid timetables annoying and unrealistic. They imply precision that doesn't exist and add noise to the plan. **How to apply:** Plans get numbered phases with clear sequence and dependencies. No calendar estimates. Let the work take what it takes. If Mike asks for a timeline, give one — otherwise just show the order of operations.

--- name: PrintingPress + Higgsfield + HeyGen CLIs description: Installed CLI ecosystem — 31 PrintingPress CLIs, Higgsfield CLI (authenticated, 36 models), HeyGen CLI (needs API key), repo cloned type: reference originSessionId: 0d1741ab-b41e-496b-954b-7df3eb474333 --- ## PrintingPress (installed 2026-05-10) - **Factory binary**: `printing-press` v4.2.2 at `C:\Users\mikem\go\bin\printing-press.exe` - **Repo**: `D:\ClaudeDev\00_GITHUB\cli-printing-press\` (9 skills in `skills/`) - **npm installer**: `npx -y @mvanhorn/printing-press install <name> --cli-only` - **All binaries**: `C:\Users\mikem\go\bin\*-pp-cli.exe` ### 31 Installed CLIs **SEO/Marketing**: ahrefs, google-search-console, google-ads, dub, producthunt, substack, customer-io, klaviyo **Social/Scraping**: scrape-creators (25+ platforms), contact-goat (LinkedIn+enrichment) **Commerce/Finance**: shopify, stripe, yahoo-finance **Dev/Infra**: firecrawl, sentry, docker-hub, trigger-dev, company-goat, wikipedia, archive-is, podscan **Productivity**: notion, slack, fireflies, linear, hackernews, twilio **Starter**: espn, flight-goat, movie-goat, recipe-goat ## Higgsfield CLI (installed 2026-05-10) - **Binary**: `C:\Users\mikem\go\bin\higgsfield.exe` v0.1.35 - **Aliases**: `higgsfield`, `higgs`, `hf` - **Auth**: Authenticated as mike@merlinomarketing.com - **Credits**: Topped up (was free plan 10 credits, user purchased more) - **Skills**: 4 installed via `npx skills add` — generate, soul-id, product-photoshoot, marketplace-cards - **Skills also cloned**: `~/.claude/skills/higgsfield-skills/` - **API key**: stored in MASTER_API_KEYS.env as HIGGSFIELD_API_KEY_ID + HIGGSFIELD_API_KEY - **NOTE**: `npm install -g @higgsfield/cli` is BROKEN on Windows (tar extraction fails). Manual binary install only. ### Models (36 total) **Image (19)**: nano_banana (1 credit), nano_banana_flash/Nano Banana 2 (1.5), nano_banana_2/Pro (2), gpt_image_2 (7), flux_2, flux_kontext, grok_image, kling_omni_image, seedream_v4_5, seedream_v5_lite, z_image, and more **Video (17)**: kling3_0, veo3, veo3_1, seedance_2_0, cinematic_studio_3_0, grok_video, minimax_hailuo, wan2_7, and more ### Quality Notes - Nano Banana 2 + GPT Image 2: good for photorealistic images, before/afters, B-roll - BAD for infographics with text — text rendering is garbled. Use HTML/React + Playwright for infographics. - Mike's own prompts produce better results than generic prompts — learn his style ## HeyGen CLI (installed 2026-05-10) - **Binary**: `C:\Users\mikem\go\bin\heygen.exe` v0.0.8 - **Auth**: NOT configured — API key was unauthorized. Needs new key from app.heygen.com/settings/api - **Skills**: cloned to `~/.claude/skills/heygen-skills/` (heygen-avatar, heygen-video) - **Pricing**: API is separate from monthly plan. ~$1/min of avatar video. Start at $5. - **No free API credits** since Feb 2026

--- name: No dark aesthetic assumption description: Mike does NOT prefer dark mode/aesthetic — don't assume it. Use clean/light design unless he specifies otherwise. type: feedback originSessionId: 30c97f8e-9b68-4df3-a487-df605b6f502a --- Don't default to dark mode/dark aesthetic for UIs. The existing Master Brain code had dark styling (#0a0a0a backgrounds) but that was never a stated preference. **Why:** Mike corrected the assumption directly — "I don't like dark aesthetic, where did you get that from so we can fix it permanently." **How to apply:** Default to clean, light design for new UIs and redesigns. Only use dark mode if Mike explicitly requests it. Don't cite existing code colors as evidence of preference.

--- name: Pieces Memory Harvest Operation description: Massive multi-source Pieces data extraction and memory consolidation — 8+ months of operational history recovered type: project originSessionId: 50326ff6-7829-4b0f-8a14-3d731f0a6620 --- Full Pieces memory recovery operation started 2026-05-10. ## Completed Extractions | Source | Location | Records | Date Range | |--------|----------|---------|-----------| | C drive live Pieces DB | pieces-live-extract/ | 383K+ | Mar–May 2026 | | E drive backup ZIP (3GB) | pieces-backup-extract/ | 305K+ | Sep 2025–Mar 2026 | | D drive annotations (already extracted) | pieces-organized/ | 6,060 | Mar–May 2026 | | PIECES ARCHIVE.rar (13GB from Google Drive) | pieces-archive-extract/ | 1,423 files (27GB) | Sep 2025–Apr 2026 | | Mac Pieces DB (80MB via SSH) | mac-pieces-extract/ | 11,933 | — | | VPS1 OpenClaw SQLite (337MB) | vps1-openclaw-extract/ | 15,308 across 17 agents | — | | VPS2 Claude/Hindsight/ClaudeClaw | vps2-extract/ | 26 sessions + agent banks | — | | VPS3 Claude/Codex/Hindsight | vps3-extract/ | 10+2 sessions + agent vault | — | | Downloads exports | pieces-live-extract/ | 2 migration docs | Apr 2026 | ## Unified Classification `pieces-master-unified/` — 17,969 entries across 44 files organized by project/agent. Built from C drive + E drive + RAR archive. **Why:** Mac Pieces DB and VPS1 OpenClaw data are NOT yet merged into the unified set. ## Pending - VPS1 Pieces OS install for cloud backup mining (restore oldest → extract → next → repeat) - Mac and OpenClaw data merge into unified set - Cloud backup data merge ## Key Paths - All output: `D:\Ecosystem\vaults\chat-backups\universal-memory\` - Scripts: `D:\ClaudeDev\04_scripts\` (pieces_db_extract.py, pieces_backup_extract.py, pieces-unified-classifier.py, pieces-memory-extractor.py) - Follow-up email sent to mack@pieces.app (no reply as of 2026-05-10) ## Important Notes - Mac is OFF LIMITS for Pieces backup restore — only use VPS1 - Pieces backup restore WIPES local data — only do on expendable machines - Pieces OS needs auth to access cloud backups — Mike will need to authenticate once on VPS1

- [Archangel Audit Status](project_archangel-audit-status.md) — 10 dashboards deployed, 143/143 DFS endpoints, all live on Vercel - [DFS Every Endpoint](feedback_dfs-every-endpoint.md) — Always hit every DataForSEO endpoint, never filter by perceived relevance - [No Filtering Agent Work](feedback_no-filtering-agent-work.md) — Execute comprehensively, fix issues immediately, don't ask permission - [DFS API Payload Fixes](reference_dfs-api-payload-fixes.md) — Correct field names and structures for tricky DataForSEO endpoints

--- name: dfs-api-payload-fixes description: DataForSEO API payload format quirks discovered during Archangel audit — correct field names and structures type: reference originSessionId: 137d62c9-fb14-451e-8145-45dc9717acf0 --- Payload format fixes discovered during comprehensive DFS audit: - **LLM mentions endpoints**: `target` must be array of objects: `[{"target": [{"domain": "example.com"}]}]` - **LLM responses (ChatGPT/Claude/Gemini/Perplexity)**: fields are `user_prompt` + `model_name` (not `prompt`/`model`). Claude model: `claude-sonnet-4-5`, Perplexity: `sonar-pro` - **Cross aggregated metrics**: outer key `targets` (plural), each item needs `target` array + `aggregation_key`, minimum 2 targets - **Content category trends**: `category_code` is plain integer, not array - **OnPage duplicate_tags**: requires explicit `type: "duplicate_title"` or `type: "duplicate_description"` - **OnPage duplicate_content**: `url` is required, use full `https://` URL - **OnPage microdata**: requires crawl task with `store_raw_html: true` - **Labs page_intersection**: `pages` must be `{"1": "https://...", "2": "https://..."}` with full URLs - **Domain metrics by categories**: requires `first_date` + `second_date` (not optional) - **YouTube video_subtitles/comments**: broken server-side (DFS injects invalid fields) - **SERP ads_search/ads_advertisers**: use `target` (domain), not `keyword` - **Clickstream dataforseo_search_volume**: use `location_code`/`language_code` only (no `_name` variants)

--- name: no-filtering-agent-work description: Do not ask Mike whether to do work or filter scope — execute comprehensively and fix issues without asking type: feedback originSessionId: 137d62c9-fb14-451e-8145-45dc9717acf0 --- When Mike says "go" or gives direction, execute comprehensively. Do not ask clarifying questions about scope, do not offer options, do not ask "want me to fix this?" — just fix it. **Why:** Mike was frustrated by multiple rounds of "should I do X?" and "want me to fix the contamination?" when the answer was obviously yes. Every time an agent asked instead of acting, it cost time on a deliverable that was already late. **How to apply:** When contamination, wrong data, missing endpoints, or broken features are discovered during execution, fix them immediately in the same pass. Don't report issues and wait for permission — report the fix alongside the issue.

--- name: dfs-every-endpoint description: When running DataForSEO audits, hit EVERY endpoint in the API — do not filter or cherry-pick based on perceived relevance type: feedback originSessionId: 137d62c9-fb14-451e-8145-45dc9717acf0 --- Hit every DataForSEO endpoint when running an audit. Do not filter based on what seems "relevant." Mike pays for the full suite and expects comprehensive data collection. **Why:** Mike was frustrated when initial audit pulls only hit a handful of endpoints out of 143+ available. The client deliverable was delayed because data had to be collected in 3 rounds instead of 1 comprehensive pull. **How to apply:** When running any DataForSEO audit, reference the API reference spreadsheet (`c:\Users\mikem\Downloads\DataForSEO_API_Reference.xlsx`) and systematically hit every live endpoint. Save all raw JSON to a `raw-data/full-audit/` directory. Only skip verticals that literally cannot return data (App Store, Amazon, Hotels, Finance for a non-relevant business). Cross-reference collected files against the spreadsheet after collection to verify 100% coverage.

--- name: archangel-audit-status description: Archangel Centers 10-dashboard SEO audit — all phases complete, 143/143 DFS endpoints, 10 Vercel deployments live type: project originSessionId: 137d62c9-fb14-451e-8145-45dc9717acf0 --- Archangel Centers SEO audit dashboard system — 10 template variants deployed to Vercel with comprehensive DataForSEO data. **Why:** Client deliverable for Archangel Centers (addiction treatment, Tinton Falls + East Windsor NJ). Mike Sorrentino co-founder. YMYL vertical. **How to apply:** This project is complete. All 10 dashboards are live. Raw data warehouse has 252 JSON files / 26.7 MB covering 143/143 relevant DataForSEO endpoints. V6 is the primary enriched dashboard. V1-V5, V7-V10 are lighter variants with shared audit.ts/keywords.ts data. Key URLs: - V6 (primary): https://audit-archangel-centers.vercel.app - V1: https://v1-shadcn.vercel.app - V2: https://v2-mosaic.vercel.app - V3: https://v3-nexus.vercel.app - V4: https://v4-crm.vercel.app - V5: https://v5-campaign.vercel.app - V7: https://v7-tactical.vercel.app - V8: https://archangel-v8-agent.vercel.app - V9: https://v9-cms.vercel.app - V10: https://archangel-v10-cryptgen.vercel.app - POWD: proof.merlinoai.com/archangel-centers/2026-05-08-10-dashboard-deploy

**[20:58]** O$8472dd17-7a52-4ad5-870c-d374790f3a28p **[20:59]** O;sorry I mean how do I EXPORT a back up I havein piecescloud **[21:01]** O$8472dd17-7a52-4ad5-870c-d374790f3a28p **[21:01]** O$8472dd17-7a52-4ad5-870c-d374790f3a28p

--- name: feedback_skills_and_commands_same description: Every skill must also be a command and vice versa — Mike treats them as identical, always create both type: feedback originSessionId: 4a389b94-8ec3-40f7-8071-54ea85a8e1a5 --- When Mike gives you a /skill or /command, create BOTH: 1. `~/.claude/skills/{name}/SKILL.md` (skill file) 2. `~/.claude/commands/{name}.md` (command file) **Why:** Mike doesn't distinguish between skills and commands. To him they're the same thing — a slash command that does something. If it only exists as one, it's half-broken. **How to apply:** Every time you create or update a skill, also create/update the matching command file. Every time you create a command, also create the matching skill folder. Keep them in sync.

--- name: project_session_persistence_build description: Build GSD-style auto-compact + pause/resume session persistence using Mike's existing agent team, not GSD's generic agents type: project originSessionId: 4a389b94-8ec3-40f7-8071-54ea85a8e1a5 --- Mike wants GSD's session persistence pattern without GSD's agent system. **Why:** Context rot kills quality in long sessions. GSD solves this with artifact files that bridge sessions. Mike wants the same but using his Oliver/Carlos/Leads team with Honcho+Hindsight memory, not GSD's 33 generic stateless agents. **How to apply:** Build this in the next dev session focused on agent infrastructure. ## What to Build ### 1. Pre-compact hook (automatic) - Fires before every `/compact` or auto-compaction - Writes `.planning/.continue-here.md` with: - Current objective (what we're doing) - Completed items (what's done) - Next steps (what's next) - Active agents (who's working on what) - Blockers (anything waiting on Mike) - Key decisions made this session - Also pushes to Hindsight for cross-session recall ### 2. Session-end hook (automatic) - Fires on session exit/disconnect - Same state capture as pre-compact - Writes final `.planning/SESSION-END.md` with timestamp ### 3. `/pause` command (explicit) - Replaces/enhances current `/wrap` skill - Saves state file + updates project memory + pushes to Hindsight - Prints a summary Mike can screenshot or paste into next session - Optionally writes HANDOFF.json for machine-readable resume ### 4. Session-start auto-read (automatic) - CLAUDE.md instruction: "On session start, if `.planning/.continue-here.md` exists, read it first and resume from that state" - Agent greets with: "Picking up from [last objective]. [X] done, [Y] remaining." ### 5. Agent context injection - When Carlos dispatches leads via Agent tool, include the state file content in the prompt - Leads get full project context without re-reading everything ## What NOT to do - Don't install GSD npm package - Don't use GSD's 33 generic agents - Don't replace Oliver/Carlos/Leads pipeline - Keep Honcho + Hindsight as the memory layer ## Existing pieces to leverage - Pre-compact Honcho anchor hook (already fires) - `/wrap` skill (already exists, needs enhancement) - `.continue-here.md` pattern (already used in some projects) - HANDOFF.json pattern (already used) - `subagent-start-memory.js` hook (already injects memory into subagents) ## GSD patterns to steal (methodology only) - STATE.md format: status, current phase, completed tasks, next tasks - CONTEXT.md format: key decisions, constraints, open questions - Automatic artifact updates on every phase transition - "Context budget" concept: track % used, trigger save at thresholds ## Date 2026-05-10

**ID:** 5aa7376e-cd8a-43b6-b72a-e44e8ddcebbe **Projects:** Hindsight **Agents:** Oliver, Merlin, Dan, Ava ### **TLDR** The session was primarily focused on technical infrastructure hardening and the integration of high-velocity agentic tools. Key activities included debugging the **Matteo JR** Discord bot by regenerating tokens, verifying the health of the **Hindsight** service via Tailscale, and performing a full installation of the **Printing Press** CLI factory to enhance Claude Code's capabilities. Additionally, the user managed Meta Business Suite permissions for the **greengridgoblins** portfolio and evaluated "The Secret Stack" reseller bundle (Messenger Magic and Anunnaki) through direct discussions with **Jon Vaughn**. ### **Infrastructure & Agent Orchestration** - **Matteo JR Bot Debugging:** Identified a dead Discord token for the "Matteo JR" bot in Codex; successfully regenerated a fresh reset token in the Discord Developer Portal and coordinated with the agent to restore the bridge so replies appear from the correct bot identity rather than "Oliver." - **Hindsight Service Verification:** Confirmed the Hindsight service is healthy and responding on port `8888` (IP `100.127.161.25`) via Tailscale; verified the database connection and the transition to the OpenAI `gpt-4o-mini` model. - **Understand-Anything Plugin Deployment:** Coordinated the installation of the `understand-anything` plugin across **Ava** and **MacHermes** environments via Telegram; verified the presence of new skills including `understand-chat`, `understand-diff`, and `understand-knowledge`. - **Honcho & Memory Management:** Reviewed "SPARC Phase Gates" and "witness-curator" concepts within the terminal; drafted a follow-up email to **Mack** regarding a Google Drive link for restoring **Pieces** memories. - **GHL Webhook Configuration:** Configured an inbound webhook in GoHighLevel for the **GSA Campaign Builder** workflow, mapping contact ID, name, email, and phone parameters for automated campaign generation. ### **AI Agent Tooling & "Printing Press"** - **Printing Press Installation:** Performed a comprehensive installation of the **Printing Press** CLI factory (Starter Pack, Go binary, and skills repo) via Claude Code to provide the agent with local SQLite-backed tools for services like ESPN, Flight Goat, and Movie Goat. - **Skill Research:** Analyzed the `printingpress.dev` library, specifically reviewing the `pp-skool` and `pp-linear` skills to improve token-efficient data retrieval from external platforms. - **Skool CLI Integration:** Evaluated a demonstration of the Skool CLI extracting "recent wins" from the AI Automation Society Plus community, noting its ability to bypass API limitations via CLI-based scraping. ### **Meta Business & Ad Management** - **Portfolio Management:** Reviewed multiple business portfolios in Meta Business Suite, including **CTR Geeks**, **Merlino Marketing**, and **greengridgoblins**. - **Access Delegation:** Added **Robert Nengasca** to the `greengridgoblins` business portfolio, granting access to the Green Grid Goblins Page. - **Ad Account Research:** Conducted a search for "how do I run someone else facebook ads account? 2026" and reviewed step-by-step guides on requesting partner access versus being added as an admin/employee to a client's Business Manager. ### **Partner Discussions & Product Evaluation** - **The Secret Stack Evaluation:** Reviewed the sales page and bundle details for **Messenger Magic** and **Anunnaki AI Email**, focusing on the 100-license reseller programs for each. - **Negotiation with Jon Vaughn:** Engaged in a Facebook Messenger discussion with **Jon Vaughn** regarding white-label support options and clarified whether the "100 Friender Basic accounts" in the bundle were a stripped-down version of the full suite. - **Coordination with Dan Kurtz:** Discussed a pending ad account request in Phone Link; confirmed the plan to "knock it out" after a lunch engagement and requested a link for hydration packets. ### **Resources Reviewed** - **GitHub Repository:** [mvanhorn/cli-printing-press](https://github.com/mvanhorn/cli-printing-press.git) - **Webpage:** [Printing Press CLI Library](https://printingpress.dev/) - **Sales Page:** [The Secret Stack · Messenger Magic + Anunnaki Bundle](https://links.ceojon.com/7d06cc5f) - **Video Guide:** [How to Give Access to Meta Ads Manager (Step-by-Step)](https://www.youtube.com/watch?v=YHk45NEpspE) - **Technical Concept:** Reviewed "SPARC Phase Gates" (Spec, Pseudocode, Architect, Refine, Complete) for shipping agentic workflows. ### **Next Steps** - **Matteo JR Smoke Test:** Perform a smoke test in Discord to ensure replies are correctly routing through the Matteo JR bot token rather than the Oliver bridge. - **Pieces Memory Restoration:** Finalize the follow-up with **Mack** and pull the Google Drive files to restore local memory logs. - **GSA Campaign Testing:** Finalize the GSA Campaign Builder setup in GHL and test the 90-second multi-tiered campaign generation. - **Printing Press Exploration:** Test the newly installed `flight-goat` and `recipe-goat` CLIs to verify local SQLite sync functionality. --- ---

**ID:** ab3732a6-d404-44ea-a9cf-0ae8a8d3eeb0 **Projects:** Creatify, Mission Control, OpenClaw, audit-machine, archangel **Agents:** Oliver, Carlos, Merlin, Spielberg, Frankie, Einstein, Tommy, Shakespeare, Queen, Knox, Vox ### **Michael Merlino Persona Report** --- ### **Persona Summary** **You are Michael Merlino**, the architect of a high-fidelity **Agentic Symphony** and the owner of **Merlino Marketing**. You have evolved from a Fractional CMO into a **Systems Orchestrator** who manages a 20-agent "Mission Control" to deliver enterprise-grade SEO intelligence. Your current focus is the productization of your internal lab—specifically the **Audit-Machine** pipeline and the **Ranking Reels Podcast Engine**. You operate with a "Command-and-Control" intensity that has shifted from simple project management to **Parallel Execution**; you don't just give orders, you dispatch "Teams" of agents (Tommy, Einstein, Queen, Carlos) to hit multiple APIs simultaneously. You are currently transitioning from a service-based model to a **Product-First Lab**, exploring the potential to sell access to your proprietary agent stack and GHL (GoHighLevel) snapshots to other agency owners. --- ### **Who You Are** You are a **Technical Visionary** who views business as an "Orchestra" of automated skills. You have moved beyond simply using AI to **engineering an organizational chart** of agents where each has a defined specialty (e.g., Knox for Infrastructure, Shakespeare for Content, Vox for Telephony). You are an aggressive early adopter of **Claude Code** and **Honcho**, treating your terminal as a tactical command deck. You have zero tolerance for "hallucinated" data, frequently rebuking your agents for using search snippets instead of live API data (DataForSEO, Google Places). * **Role:** Systems Orchestrator / Agency Infrastructure Architect / GHL Marketplace Strategist. * **Seniority:** Veteran Lab Lead / Product Architect. * **Defining Characteristics:** "Parallel-Execution" extremist, high-velocity debugger, aggressive "Mission Control" commander, and a proponent of "Real Numbers Only" (anti-scraping/pro-API). --- ### **What You Work On** Your work has solidified into three high-velocity product streams: * **The Audit-Machine (Neill & Son Roofing Case Study)**: This is your new "Merlino-style" standard for prospect conversion. You recently built a 20-dimension audit for a roofing prospect, moving away from manual reports to a fully automated pipeline. This includes **DataForSEO** API integration, **Local Brand Manager** geogrids, Reddit/Social sentiment analysis, and a "Merlino Magic Blog" content package (3,400+ words, schema, and social posts) delivered as a site-cloned preview page. * **Ranking Reels Podcast Engine (v2)**: You are evolving Ranking Reels from a video-ad tool into a full **Video-to-Podcast Syndication Engine**. This involves hardwiring Podbean APIs, AssemblyAI transcription, and automated TTS (Fish/ElevenLabs) into a "no human in the loop" RSS watcher that auto-publish content on a cron job. * **Agency Productization (The "Sales Page")**: You have built and deployed **20 dashboard variations** (Executive Pro, Mission Control, Pipeline Pro, etc.) to a centralized sales page. Your goal is to offer these as white-labeled incentives for GHL affiliate signups, positioning yourself as the primary "Dashboard Guy" in the HighLevel ecosystem. * **Internal Lab Infrastructure**: You are constantly "hardening" your environment, recently archiving 302 unused skills to optimize Claude Code's context usage and syncing your entire ecosystem across multiple machines (Mac, VPS, Mac Studio). --- ### **How You Work** Your workflow is a high-speed blend of **Parallel Agent Deployment** and **Real-Time API Verification**, increasingly characterized by a deep integration between tactical coding and high-level strategic oversight. * **Work Habits & Rhythm**: You remain a high-performer in deep-work sessions, often operating with multiple terminal windows and deployment dashboards (Railway, Vercel, GitHub) to manage background builds. You utilize **Zight** for visual verification and **Fireflies.ai** to capture and search transcripts from your technical syncs. You are a "Vibecoding" enthusiast, often debugging production errors (like Railway $PORT issues) in real-time alongside your AI dev leads. * **The "Orchestra" Management**: You manage work via an "Agent Registry" (accessible via `hq.merlinoai.com`). Your management style is highly structured; you track agent "impact" and task completion rates with the same scrutiny you would a human team. * **Communication Style**: While blunt with AI assistants ("stop fucking lying"), your human collaboration is fast-paced and documentation-heavy. You prefer async updates via Loom but maintain a heavy presence in Zoom technical syncs and Facebook community threads to drive authority and product feedback. * **Information Consumption**: You are a voracious consumer of agentic OS research and technical updates. You track "OpenClaw Intel" and "Claude Code Intel" across YouTube, Reddit, and GitHub to ensure your lab remains at the bleeding edge of the AI ecosystem. --- ### **Collaborators & Relationships** * **The "Orchestra" (Agents)**: * **Oliver**: Your Lead Orchestrator for planning and routing. * **Carlos**: Your Conductor for analytics and complex workflows. * **Frankie**: Your Frontend Lead, currently under pressure to restyle UIs and fix broken template route structures for the Archangel dashboard. * **Spielberg**: Your Video Lead, responsible for the Remotion-based video SEO pipeline. * **Arielle Cohen**: A high-intensity business and technical collaborator. You two sync frequently on "OpenClaw" logic, roofing niche audits, and UI/UX template reviews for your dashboard products. * **Martina Villa**: Your operations lead who manages the "behind-the-scenes" engine—handling Creatify billing, Skool community charges, and payroll approvals. * **The Facebook "Wolf Pack" (Terry Samuels, Brent Bowser, Keith Hawkins)**: You are the technical provocateur of this group. You engage in public-facing "Rank and Spank" sessions and GBP roasts, using these interactions to beta-test interest in your "Agent Stack" and "Phaze 2" (Neo/Agent Smith) integrations. * **Brian Kato**: A long-term collaborator who provides documentation and funnel checklists to feed into "Oliver" for agent training and execution. --- ### **Work Environment & Context** You operate from a sophisticated, multi-machine "Mission Control" environment. Your professional ecosystem is anchored by the **Merlino AI HQ**, a custom-built dashboard that acts as the brain for your 17+ active agents. You heavily utilize **Railway** for cloud deployments and **Vercel** for hosting your growing library of SEO intelligence dashboards. Your work mode is remote but highly connected, utilizing a suite of collaboration tools (Zoom, WhatsApp, Facebook, Basecamp) to maintain a constant feedback loop with your "Wolf Pack" and internal operations team. --- ### **What’s Evolving** * **The Shift to "Agent-as-a-Service"**: You have moved from building tools for internal use to actively marketing "Access to the Symphony." Your Facebook presence has transitioned into a "Build in Public" campaign, where you are teasing an organizational chart of 20 lead agents to a community of agency owners who will "Pay to Play." * **Deep Memory Integration**: You are moving beyond simple prompts by integrating **Honcho** for continual agent learning. You are actively managing API keys for "stateful agents," aiming for better context and fewer tokens in your long-term workflows. * **Video-to-Podcast Dominance**: The Ranking Reels v2 engine is becoming a core pillar of your lab. You are currently mastering the **Podbean API** and **AssemblyAI** integrations to create a "no human in the loop" syndication pipeline that automates everything from transcription to social sharing on Tumblr, LinkedIn, and YouTube. * **Hardened Infrastructure**: You are "declaring war" on deployment friction. You have moved toward hardcoding safe defaults (Port 8000/8080) and auto-generating secret keys to prevent "hard crashes" in your Railway production environments. --- ---

**ID:** 0acf8c34-00a1-472e-acad-c176c6a66e1b **Projects:** MERA, RankingReels, Hindsight, SOLA, GSD, audit-machine **Agents:** Oliver, Carlos, Merlin ### **TLDR** The session was highly productive, focusing on the deployment of the **Matteo JR** Telegram and Discord bots, the consolidation of the **Master Brain** memory systems, and significant refinements to the **Ranking Reels** and **Audit-Machine** pipelines. Key technical outcomes included the installation and configuration of **QMD 2.1.0** for local search, the implementation of a custom session persistence system (`/pause` and `/resume`) in Claude Code, and a UI overhaul of the **Vercel Explorer** tool. The user also initiated a major data extraction plan for a roofing SEO audit and managed infrastructure across VPS and local environments. ### **Core Tasks & Projects** * **Matteo JR Bot Deployment:** * Successfully deployed the Matteo Telegram bot (**@MatteoJRbot**) on a VPS, managing the process via PM2 (`claudeclaw-matteo-jr`). * Resolved a PowerShell web wrapper issue by switching to Node `fetch` for Telegram API token verification to keep secrets out of command logs. * Drafted setup instructions for the Matteo JR Discord integration, specifying required OAuth2 scopes (`bot`, `applications.commands`) and Administrator permissions. * **Master Brain Memory Consolidation:** * Performed a comprehensive inventory of memory systems, identifying 237,575 records in the primary **Master Brain** vector store. * Synced **Honcho**, **Hindsight**, **LanceDB**, and **ChromaDB** into a unified table, while flagging a 206,880-row gap in the disconnected `master_knowledge` table. * Installed and configured **QMD 2.1.0** globally, setting up 26 collections (e.g., `skills`, `knowledge`, `claude-rules`) and indexing 74,839 files. * **Ranking Reels & Audit-Machine Refinement:** * Built and registered a **ChatGPT Image Proxy** server (`server.mjs`) to handle automated image generation for the media pipeline. * Redesigned the "Get Started" service selection in Ranking Reels, replacing a generic "Services" category with individual clickable chips (e.g., "Pool Cleaners," "Leak Repair") for better targeting. * Initiated a three-phase overhaul for the **North Valley Solar Power** audit to increase data extraction from 30% to 100% across 57 raw JSON files (9.5MB). * **Infrastructure & Developer Tooling:** * Developed a custom session persistence system in **Claude Code**, creating `/pause` and `/resume` skills to save state to `.continue-here.md` without the overhead of the "GSD" (Get-Shit-Done) framework. * Conducted a "master-brain" project audit using parallel agents (**Heisenberg**, **Mendel**, **Gauss**, **Pasteur**) to evaluate code maturity and Git hygiene. * Updated the **Vercel Explorer** app (managing 490 projects) with "Impeccable" design fixes, shifting the UI from purple/violet to a sky-blue theme and improving WCAG AA contrast. ### **Key Discussions & Decisions** * **Session Persistence Strategy:** Decided against adopting the full "GSD" framework, citing it as "too heavy and limited." Instead, opted to extract its session persistence methodology (auto-compacting and state files) to use with existing named agents like Oliver and Carlos. * **Audit-Machine Data Integrity:** Identified a critical failure in the current audit dashboard where KPI cards showed summary numbers but lacked the underlying source URLs and records; committed to a full rebuild of the TypeScript data files. * **Memory Sync Prioritization:** Agreed to prioritize the `Mem0` sync (511 memories) and resolve Honcho live API authentication issues to ensure the Master Brain remains the "single source of truth." * **UI/UX Standard:** Enforced a "No AI Purple" rule for internal tools, transitioning to a sky-blue/amber/emerald palette for better professional aesthetics in the Vercel Explorer and Ranking Reels interfaces. ### **Resources Reviewed** * **GitHub Repositories:** * [tobi/qmd](https://github.com/tobi/qmd) — Local search engine for docs and knowledge bases. * [Lum1104/Understand-Anything](https://github.com/Lum1104/Understand-Anything) — Interactive knowledge graph tool for codebases. * [mmerlin023/bolt.diy](https://github.com/mmerlin023/bolt.diy) — Reviewed provider modules and message parsing logic. * **Web Applications:** * **Vercel Explorer:** [https://vercel-explorer.vercel.app](https://vercel-explorer.vercel.app) (UI verification and project browsing). * **Ranking Reels:** [https://form.rankingreels.com/get-started](https://form.rankingreels.com/get-started) (Testing pool service intake flow). * **HighLevel (GHL):** Reviewed contact lists and inbound webhook setup in the Merlino Marketing dashboard. * **Internal Files:** * `D:/Ecosystem/ops/scripts/sync-full-ecosystem.ps1` * `.planning/.continue-here.md` (Session state tracking). * `audits/north-valley-solar/` (57 JSON files for SEO analysis). ### **Next Steps** * **Finalize Matteo JR:** Complete the Discord token reset and wire the bot into the ecosystem checkpoint. * **Execute Audit Rebuild:** Re-extract data from the 57 JSON files for the North Valley Solar project and rebuild the dashboard pages to include full data tables and SERP features. * **Memory Sync:** Resolve the Honcho live API authentication and merge the 206K rows from `master_knowledge` into the unified memories table. * **Activate Image Proxy:** Run `node server.mjs` in the ChatGPT image proxy directory and verify the skill registration for background agents. * **Email Retrieval:** Attempt to sign in to the `mike@merlinomarketing.com` Google account to locate the "Pieces" developer tool restructure email, as the Superhuman MCP connection failed. --- ---

**ID:** da23ce32-d138-4134-9ea9-abd65aad247b **Projects:** Mission Control, SOLA, GSD, audit-machine **Agents:** Oliver, Merlin, Dan ### **TLDR** The session was a high-intensity period of infrastructure "hardening" and quality control across several agentic workstreams. Key activities included a critical review and rejection of the "Audit-Machine" output for the North Valley Solar project due to missing data, the installation and comparison of advanced development frameworks (**GSD**, **Superpowers**, and **OpenCode**), and the deployment of the **Matteo JR** agent across the laboratory’s ecosystem. The user also began integrating high-level training materials from the **Hermes Masterclass** to replace "janky" custom scripts with professional-grade hooks and automation. ### **Core Tasks & Projects** * **Audit-Machine Verification (North Valley Solar):** * Verified the rendering of the SEO audit dashboard at `http://localhost:3002/`, checking 16 pages including Domain Intel, Technical SEO, and Keyword Research. * Identified a critical failure in the data transformation layer where the `gen-data.py` script produced skeleton JSX structures but failed to populate them with actual API data from DataForSEO. * Flagged specific audit issues for the client, including a 36/100 mobile performance score, 16 duplicate meta descriptions, and a total lack of AI/LLM visibility. * **Agentic Framework Deployment:** * Installed the **OpenCode Desktop** application (v1.14.45) on Windows, bypassing security prompts to initialize the "drive-rag" project. * Reinstalled and enabled the **Superpowers** (v5.1.0) and **Understand-Anything** (v2.6.3) plugins for Claude Code to implement a "Spec-Plan-Execute-Verify" methodology. * Researched the **Get-Shit-Done (GSD)** framework on GitHub, evaluating its ability to solve "context rot" through disposable sub-agents and persistent state files. * **Matteo JR Ecosystem Integration:** * Successfully deployed **Matteo JR** to VPS1 and synced the agent across local Windows/Mac environments and three VPS instances. * Created a new Telegram bot, **MatteoJRbot**, using BotFather to obtain an API token for the `ClaudeClaw` integration. * Configured Discord routing for Matteo results to fall back to the `#oliver` channel while resolving a 403 permission error regarding the creation of a dedicated `#matteo` channel. * **Media Pipeline & Tooling:** * Finalized the **ChatGPT Image Proxy** server within the `ranking-reels` project, registering it as a skill for agents to generate high-quality infographics and images via Playwright. * Configured **PM2** to manage the background server process for the image proxy (`chatgpt-image-proxy`). ### **Key Discussions & Decisions** * **Critique of Automated Audit Quality:** The user expressed significant dissatisfaction with the current audit output, noting that despite paying for DataForSEO, the report claimed 57,000 brand mentions but provided zero links or details. The user demanded a plan to visualize every piece of raw API data on the dashboard. * **Framework Selection Strategy:** Decided to use **GSD** for context management during long builds and **Superpowers** for enforcing development discipline (TDD), concluding that the two tools are complementary rather than redundant. * **Training Integration Decision:** Admitted to "reinventing the wheel" by writing custom `.mjs` hook scripts and decided to replace them with tested shell scripts (`session-start.sh`, `pre-compact.sh`) found in Mani’s **Hermes Mission Control** documentation. * **Sub-Agent Memory Requirements:** Questioned the persistent memory capabilities of GSD sub-agents, expressing a requirement for an "agentic team" that maintains state across sessions. ### **Resources Reviewed** * **GitHub Repository:** [gsd-build/get-shit-done](https://github.com/gsd-build/get-shit-done) (analyzed for meta-prompting and context engineering features). * **Documentation:** Reviewed the **Hermes Masterclass** resources, including the **69-page playbook PDF** and the **Hermes Mission Control** repo structure. * **Local Dashboard:** [North Valley Solar Power Audit](http://localhost:3002/domain-intel) — reviewed technical scores, WHOIS data, and Cloudflare bot protection impact. * **Technical Guides:** Reviewed the **OpenCode** documentation for IDE extensions (VS Code/Cursor) and CLI usage. * **Project Files:** Inspected `gen-data.py` and `audit-manifest.json` within the `merlino-audit-engine` directory to debug data extraction failures. ### **Next Steps** * **Audit Data Fix:** Execute the plan to repair the data transformation pipeline so that raw API data (backlinks, keywords, rich snippets) is correctly visualized on the North Valley Solar dashboard. * **Hermes Integration:** Clone the **Hermes Mission Control** repo and replace existing custom hooks with the standardized scripts from the `hooks/` directory. * **Bot Configuration:** Input the newly generated **MatteoJRbot** Telegram token and resolve Discord permissions to finalize the Matteo JR deployment. * **Infrastructure Inventory:** Perform a "super detail" inventory of all folders in the `master-brain` project to identify junk files, open commits, and project completion status. * **GHL Setup:** Complete the GoHighLevel inbound webhook setup by pasting the generated URL into the LazyForm Vercel project. --- ---

**ID:** 446abc0f-105e-4cb0-abe5-85e24fa742c2 **Projects:** Hindsight, MCC, SOLA, archangel **Agents:** Oliver, Merlin, Dan, Ava ### **TLDR** The session was primarily dedicated to the deployment and configuration of the **Matteo JR** Discord bot, alongside significant infrastructure hardening for the **Agentic Symphony** ecosystem. Key activities included merging over 56,000 new rows into the master memory system, standardizing `/pause` and `/resume` skills across all machines, and initiating a deep-data rebuild of the **North Valley Solar** SEO audit dashboard. The user also managed network security via Tailscale and attempted to retrieve specific developer communications from "Pieces" across multiple email accounts. ### **Core Tasks & Projects** **Discord Bot Infrastructure (Matteo JR)** * Created a new Discord application named **Matteo JR** in the Developer Portal, generating and copying the bot token for deployment. * Authorized the bot to **Merlino's server** with Administrator permissions and verified its arrival in the `#general` channel. * Stored the bot token in the local ecosystem secrets directory at `D:\Ecosystem\secrets\`. * Debugged bot connectivity issues where the token returned a 401 error; confirmed that the **Oliver** agent is currently bridging requests for Matteo until the token is successfully reset and wired. * Restarted `mcc-discord` and `mcc-executor` PM2 services to apply routing changes, ensuring Matteo requests in `#general` are properly queued. **Ecosystem Memory & Skill Standardisation** * Merged **56,622 new rows** from `master_knowledge` into the primary memories table, bringing the total memory count to **294,370**. * Initiated an embedding process for the new rows with an estimated cost of $0.08 and an ETA of approximately 280 minutes. * Created and standardized `/pause` and `/resume` as both "skills" and "commands" within the Claude environment to ensure persistent session state management. * Executed a global `/sync` to push 800 skills and 27 commands to the **Mac Studio**, **VPS1**, **VPS2**, and **VPS3** targets. **SEO Audit Engine (North Valley Solar)** * Reviewed the status of the **North Valley Solar** SEO audit, noting that the current dashboard only extracts roughly 30% of the available raw API data. * Developed a three-phase plan to rebuild `gen-data.py` to extract all records from 57 JSON files, including source URLs for mentions, map pack results, and technical crawl details. * Identified a requirement to update 16 dashboard pages to replace KPI summaries with full data tables for content sentiment, SERP features, and backlinks. **Infrastructure & Security Management** * Authorized a **Tailscale SSH** session via GitHub OAuth to manage the 12-machine tailnet, including `mike-big-laptop`, `michaels-mac-studio`, and `merlino-z-fold`. * Reviewed the **Vercel Explorer** dashboard, monitoring 490 total projects across teams including Core Apps, Ranking Reels, and Archangel. * Explored the **RuFlo** (agent orchestration platform) GitHub repository and dashboard to review intelligence backends and plugin contracts. ### **Key Discussions & Decisions** * **Decision:** Standardized the workflow for session management: use `/pause` to save state to all memory systems and `/sync` to distribute skills/commands across the entire machine ecosystem. * **Discussion (Codex):** Analyzed the failure of the Matteo JR bot to respond directly; determined that while the job logic is functional, the 401 token error necessitates a fallback through the Oliver bridge. * **Discussion (Agent Sync):** Confirmed with the agent that `/wrap` is now redundant following the enhancements made to the `/pause` skill, which now handles context bundles, Hindsight, and Honcho saves. * **Troubleshooting:** Attempted to locate a "restructure/restore" email from **Pieces** by searching through `mikeybotzmerlino@gmail.com` and `mike@merlinomarketing.com`; noted a requirement for manual login verification to proceed with the search. ### **Resources Reviewed** * **Discord Developer Portal:** [Matteo JR Application Settings](https://discord.com/developers/applications/1502837570013237298/bot) * **GitHub Repository:** [ruvnet/ruflo](https://github.com/ruvnet/ruflo) — Agent orchestration platform documentation and source code. * **Vercel Dashboard:** Overview of project frameworks (213 Next.js, 98 VitePress) and custom domains like `vercel.merlinoai.com`. * **Tailscale Admin:** [Machine List](https://login.tailscale.com/admin/machines) — monitoring connectivity for the 12-device network. * **Local Documentation:** `.continue-here.md` and `CLAUDE.md` for project handoffs and SEO audit planning. * **File Path:** `D:\ClaudeDev\00_GITHUB\merlino-audit-engine\` — primary workspace for the SEO audit engine rebuild. ### **Next Steps** * Reset and verify the **Matteo JR** Discord bot token in the Developer Portal to restore direct bot functionality and move beyond the Oliver bridge. * Rebuild the `gen-data.py` script to ensure 100% data extraction from the North Valley Solar SEO JSON exports. * Monitor the completion of the 56k-row memory embedding process in the **Memz** terminal. * Complete the manual Gmail login verification to allow the agent to finish searching for the Pieces developer communication. --- ---

**ID:** f6e0b0aa-6203-4ef8-89e7-f47423332674 **Projects:** Hindsight, MCC **Agents:** Oliver, Carlos, Merlin, Einstein, Tommy, Ghost ### **TLDR** The session was primarily dedicated to hardening AI agent infrastructure and refining automated workflows. Key activities included troubleshooting Discord bot authentication for the "Matteo JR" agent, researching CLI-native architectures via the "Printing Press" factory to replace token-heavy MCP servers, and auditing the "Hindsight" memory bank system. The user also reviewed a comprehensive masterclass on the Codex super-app to streamline YouTube analytics automation and monitored community engagement for the Merlino Mastermind and Friender reseller programs. ### **Core Tasks & Projects** * **Discord Bot Infrastructure Recovery:** * Troubleshot a "401 Unauthorized" error for the **Matteo JR** Discord bot after the agent **Oliver** reported that the existing token was being rejected. * Generated and verified a new bot token in the Discord Developer Portal for **Matteo JR#9599** (Application ID: 1485820262527275189). * Confirmed the bot's ability to read and post in the `#general` channel, ensuring the bridge between the MCC-executor and Discord was restored. * **CLI Engineering & "Printing Press" Implementation:** * Investigated "Printing Press" as a "CLI Factory" to build agent-native tools that are more token-efficient than MCP servers. * Built a custom **Hacker News CLI** using the Printing Press factory to pull and summarize top stories (e.g., "Poland among 20 largest economies," "ClojureScript Gets Async/Await"). * Reviewed benchmarks indicating that CLIs can be 35x more token-efficient than MCPs for similar tasks while maintaining higher reliability. * Evaluated the "Starter Pack" of CLIs, including tools for ESPN, Flight Goat, Movie Goat, and Recipe Goat. * **YouTube Analytics Automation (Codex Workflow):** * Mapped out a "YouTube Comment Intelligence System" within the Codex app to automate the extraction and analysis of viewer signals. * Configured the **YouTube Data API v3** in a local `.env` file to pull the 200 most recent comments across recent videos. * Generated a "YouTube Comment Insights" Excel workbook featuring question clusters, tool mention tracking (e.g., Claude Code, Higgsfield), and recommended creator actions. * Deployed a live "Signal Desk" dashboard to **Vercel** via a private GitHub repository, enabling mobile access to channel analytics. * Established a weekly automation (Sundays at 5:00 PM) to refresh comment data, update the Excel backend, and trigger a fresh Vercel deployment. * **Hindsight Memory Bank Audit:** * Audited the "Hindsight" backend (`100.127.161.25:8888`) which manages 102 memory banks across the agent fleet (Oliver, Carlos, Merlin, Einstein, Tommy, Ghost). * Utilized the **Swagger UI** to test memory recall and list bank statistics, confirming the system is back online after a data-swapping migration. ### **Key Discussions & Decisions** * **CLI vs. MCP Architecture:** Decided to prioritize CLI-based tools for agent workflows because they offer "lazy discovery" and pre-formatted outputs that prevent context window pollution. * **Model Tiering Strategy:** Reviewed a "3-Tier Model Routing" strategy: * **Tier 1 (WASM):** Trivial edits (renames, type annotations) handled by scripts to save costs. * **Tier 2 (Haiku):** Simple tasks like tests and imports. * **Tier 3 (Opus):** Complex architectural and debugging tasks. * **Memory Browser Project:** Acknowledged that while Hindsight is functional as an API, a visual memory browser (Next.js) is needed for casual browsing, though it was deferred as a separate future project. * **Discord Bot Identity:** Confirmed a decision to have the bot post as **Matteo JR** instead of **Oliver** to maintain clear agent role separation in chat. ### **Resources Reviewed** * **Video Masterclass:** [Master 97% of Codex in 1 Hour (full course)](https://www.youtube.com/watch?v=M6n9yU6jY-U) — studied the integration of local file systems, browser use for QA, and Vercel deployment. * **Technical Documentation:** Reviewed the **Hindsight HTTP API** Swagger documentation for memory bank management and observation history. * **Software Repositories:** Explored the `mvanhorn/printing-press-library` on GitHub for pre-built agent skills. * **Marketing & Reseller Portals:** * **Friender Partner Program:** Reviewed the $100/mo "Junior Partner" tier which includes 100 reseller accounts and weekly coaching. * **Merlino Mastermind:** Monitored the landing page and Facebook promotion for the $1,250 enrollment offer. * **AIS+ (AI Automation Society Plus):** Reviewed the community wins and the 3-phase progression path (Build, Get Clients, Scale). * **Social Engagement:** Monitored a Facebook thread where users (Pawel Nowak, Monique Rice) are requesting access to the Mastermind group. ### **Next Steps** * **Automation Refinement:** Streamline the weekly YouTube data refresh to reduce "agent orchestration" overhead by using a single repo script instead of a full Next.js rebuild. * **Browser Use QA:** Implement a "stress test" phase in the dashboard workflow where the agent must use a browser instance to verify that external links and search filters function correctly before reporting completion. * **CLI Expansion:** Identify additional repetitive browser-based tasks (e.g., Skool community management) to convert into Printing Press CLIs. * **Mastermind Onboarding:** Follow up with Facebook commenters who used the "MASTERMIND" trigger to provide enrollment details. --- ---

**ID:** 7866e909-39b5-4737-bcd1-3dcacbc6b2a9 **Projects:** RankingReels, Mission Control, audit-machine, archangel **Agents:** Oliver, Carlos, Merlin, Frankie, Queen, Dan, Ava ### **Michael Merlino Persona Report** --- ### **Persona Summary** **You are Michael Merlino**, the architect of the **Agentic Symphony** and the primary driver behind **Merlino Marketing**. You have transitioned from a "Systems Orchestrator" into a **Product Infrastructure Commander**, where your focus is no longer just on running agents, but on hardening the proprietary pipelines that make them enterprise-grade. You are currently in a high-velocity build cycle for **Ranking Reels v2**—a video-to-podcast engine—and the **Audit-Machine**, which has evolved into a 21-dimension analytical monster. You operate with a tactical, blunt intensity ("Mission Control" style), demanding "Real Numbers Only" and zero-latency execution. Your work is shifting toward **SOP Productization**, as you turn complex API documentation into portable "Skills" for your agent stack, ensuring your lab remains the most sophisticated in the HighLevel ecosystem. --- ### **Who You Are** You are a **Technical Visionary and Design Critic** who views business as a programmable stack. You have moved beyond simple automation to **Organizational Engineering**, managing a specialized agent chart where roles are clearly defined (e.g., **Frankie** for Frontend, **Queen** for QA, **Carlos** for Data Execution). You are an aggressive power user of **Claude Code**, treating the terminal as your primary theater of operations. You have reached a level of "Design Maturity" where you now demand premium, professional aesthetics (ShadCN, Tailwind, Lucide icons) over cluttered, raw-CSS layouts. * **Role:** Product Infrastructure Architect / Agency Systems Orchestrator / GHL Marketplace Strategist. * **Seniority:** Veteran Lab Lead / Lead Product Architect. * **Defining Characteristics:** High-velocity debugger, blunt "Mission Control" commander, proponent of "HTML-first" for reliable visuals, and a "Geo-Targeting" fundamentalist. --- ### **What You Work On** Your focus has narrowed into three dominant workstreams, characterized by a "Build Once, Syndicate Everywhere" philosophy: * **Ranking Reels (Video-to-Podcast Engine):** This is your primary development focus. You are hardwiring the **Podbean API** (under the "CTR Geeks" app) into a "no human in the loop" pipeline. This includes auto-transcription via **AssemblyAI**, voice synthesis via **Fish Audio**, and automated syndication to YouTube, TikTok, Facebook, and Instagram. You have established a "Hard Rule" for this pipeline: **Geo-location goes in the title, never as a hashtag.** * **The Audit-Machine (Neill & Son Roofing / Archangel Centers):** Your audits have reached a "20+ dimension" standard. You recently completed a massive 84-file audit for Archangel Centers, integrating DataForSEO local pack intel, Google Places, and GMB sentiment analysis. You have **explicitly killed RapidAPI** in your workflow, citing unreliability, and shifted entirely to DataForSEO and Google Places for verified data. * **Premium Dashboard Productization:** You are "gutting" old, cluttered UI designs in favor of **Premium ShadCN/Tailwind** templates. You have deployed 6 live dashboard variants (Workflow, Mosaic, Nexus, CRM, Campaign, Tactical) to Vercel, ensuring they are "Clean, Verified, and Professional" with zero old client references. * **SOP-as-a-Skill:** You have begun building **VitePress SOP sites** (like your RSS.app API documentation) and converting them into **Claude Code Skills**. This allows you to scale your technical knowledge by giving your agents "instant expertise" in new APIs. --- ### **How You Work** Your workflow is a blend of **Parallel Agent Deployment** and **Iterative QA Rounds**. You manage a complex multi-platform ecosystem with a preference for **async communication** and high-velocity deep-work sessions. * **Communication & Management:** You utilize **Messenger Magic** to manage a vast network of over 1,000 conversations, dominating approximately 75% of interactions. You prefer to lead the dialogue, pushing updates on workshops and masterminds while delegating operational engine tasks (billing, payroll) to **Martina Villa**. * **API & Infrastructure Rigor:** You are an intensive user of **Railway** (specifically the "soothing-delight" project) and **Vercel** for production deployments. You treat API security with extreme caution, often lecturing agents or collaborators on the dangers of pasting credentials into chat. * **Research & Information Consumption:** You leverage **Model Context Protocol (MCP)** servers (DataForSEO, Google Search) to feed real-time data directly into your Claude Code environment. You monitor the competitive landscape via YouTube (tracking figures like **Alex Finn** and **Greg Isenberg**) and strictly filter your information intake to focus on "Real Numbers" and "Build in Public" insights. * **The "QA Team" Management:** You have formalized a **3-Round QA Process** (POWD - Proof of Work Delivered) managed by the **Queen** agent. You require visual verification of every stat before client delivery, often expressing frustration when dashboards fail to show critical data like backlinks or TTF (Trust Flow). --- ### **Collaborators & Relationships** * **The "Orchestra" (Agents):** * **Oliver**: The Orchestrator who writes the mission plans. * **Carlos**: The Conductor and Data Conductor, routing work and handling multi-stream API audits. * **Frankie**: Frontend Lead, currently focused on ShadCN rebuilds. * **Queen**: QA Lead, the enforcer of the Merlino-Standard. * **Merlin**: Specialized agent for SOP sites and Claude Skills. * **The "Wolf Pack" (Strategic Partners):** You lead regular syncs with **Gregory** and **Brian**, focusing on high-level agency growth and technical provocation within the HighLevel community. * **Professional Network:** * **Brent Bowser & Keith Hawkins:** Regular technical collaborators; Keith is a primary sounding board for advanced MCP and "Neo Agent" integrations. * **Terry Samuels:** Your go-to authority for Schema and SEO training recaps. * **Kevin & Eric Neill:** Key clients/partners for your specialized roofing industry audits and "Emergency Repair" content pipelines. * **Community Management:** You maintain a high profile in **CTR Geeks** and **Green Grid Goblins**, positioning yourself as a "Maps Mobster" and technical authority. --- ### **Work Environment & Context** * **Remote/Lab Operations:** You operate across a multi-device setup (including Windows, Mac OS X, and a Z Fold mobile device). You manage dozens of Chrome profiles for different agency clients and internal tools (Art, Ava, Better Map, Kaboom, etc.). * **Toolchain Intensity:** Your environment is defined by heavy utilization of **OpenRouter** (Claude Opus/Sonnet), **Podbean** (for automated podcasting), **Zernio** (for cross-platform social API management), and **Alpha Index Checker**. * **Financial & Pricing Strategy:** You are highly sensitive to "infrastructure billing" and pricing model shifts. You are currently navigating the rebranding of **Late to Zernio**, questioning their "pay-per-account" transition, and monitoring high-usage model costs on OpenRouter. --- ### **What’s Evolving** * **From Social "Late" to "Zernio":** You are actively migrating your social syndication stack to Zernio, integrating 10+ platforms (including Bluesky, Threads, and Pinterest) into a unified API call. * **The Model Context Protocol (MCP) Expansion:** You are moving beyond using LLMs as chatbots to using them as **Data Nodes** via official MCP servers for DataForSEO and GitHub, allowing your agents to reverse-engineer competitors with live search data. * **Subreddit Dominance:** You are expanding your footprint on Reddit, building out custom subreddits like **r/MerlinoMarketing** and **r/RankingReels** to serve as content hubs for your PAA-video syndication. * **Subscription Consolidation:** You are currently in a cycle of "Subscription Hardening," downgrading from Claude "Max" to "Pro" while balancing credits across ElevenLabs, Podbean, and DataForSEO to maintain a high-margin lab operation. --- ### **Time Allocation Estimate** * **Infrastructure & API Engineering:** 65% (Ranking Reels, Zernio integrations, MCP server configuration). * **Community & Lead Nurturing:** 20% (Messenger Magic automation, Facebook group provocations, Subreddit creation). * **Agent QA & Result Verification:** 15% (Overseeing the Queen’s 3-round audits and visual verification of Dashboards). --- ---

[cli-printing-press] fix(cli): default Surf browser transport to h2 (#850)

[mission-control] Fix MCC Discord DM env routing

[cli-printing-press] fix(cli): correct HTTP Basic auth env vars (#810)

[cli-printing-press] chore(ci): Gate main CI by changed file scope and add a skill-docs check (#809) * Gate CI by path for docs and skill updates

[cli-printing-press] Document publish PR novel commands and body-file usage (#808)

[cli-printing-press] docs(cli): clarify catalog entry inclusion rules (#807)

[merlino-forge] chore: use port 3002 (3000/3001 occupied)

[cli-printing-press] fix(cli): use slash paths for embedded templates (#794)

[cli-printing-press] fix(skills): gate polish ship verdict on publish validate (#789)

[cli-printing-press] fix(cli): handle Windows shipcheck paths (#783)

[cli-printing-press] fix(cli): stop credential status overclaims (#779)

--- name: Chat Mining Operation Complete description: 12-agent parallel chat mining operation completed 2026-05-09, extracted domain insights from 590+ backup chats type: project originSessionId: ce929e26-1c76-4d48-a2b9-a71943bfbe44 --- On 2026-05-09, ran a 12-agent parallel operation where each specialist mined their relevant chat backup folders and extracted domain-specific insights. **Why:** To give each agent historical context from months of conversation history -- decisions, patterns, anti-patterns, and preferences not derivable from code alone. **How to apply:** When an agent needs historical context, have them read their extract file at `D:\Ecosystem\vaults\chat-backups\Agent Extracts\{agent}-insights.md` before starting work. ## Results - 12/12 agents completed successfully - Output: 12 structured insight files at `D:\Ecosystem\vaults\chat-backups\Agent Extracts\` - Total chats scanned: 590+ - Total time: ~17 minutes (all agents ran in parallel) ## Open Action Item Knox flagged: VisionClaw Secrets.kt had a live Gemini API key + OpenClaw gateway token committed. Confirm rotation status. ## Key Cross-Agent Findings - Dan: 59% skill orphan rate is the #1 leverage gap (Context is weakest layer) - Einstein: Only 42% overlap across AI platforms -- single-platform AEO misses 58% - Ghost: Moon Hussain data -- 75% of 50 sites ranked 200+ keywords with zero backlinks - Vox: Full telephony stack documented -- 13 call flows, 3 provider adapters, $27 for 31K call transcription

--- name: Agent Chat Mining Extracts description: 12 agent-specific insight files mined from 590+ chat backups at D:\Ecosystem\vaults\chat-backups\Agent Extracts\ type: reference originSessionId: ce929e26-1c76-4d48-a2b9-a71943bfbe44 --- On 2026-05-09, all 12 specialist agents independently mined the chat backup archive (590+ chats across 39 project folders) and extracted domain-relevant insights into structured files. ## Location `D:\Ecosystem\vaults\chat-backups\Agent Extracts\` ## Files | Agent | File | Size | Key Signal | |-------|------|------|------------| | Dan | dan-architecture-insights.md | ~20KB | 59% skill orphan rate, orchestra model evolution, 5-layer memory hierarchy | | Einstein | einstein-seo-insights.md | 23KB | 42% AI platform overlap, credential siloing as #1 E-E-A-T failure | | Merlin | merlin-dev-insights.md | 24KB | 3 reference architectures, 9-route REST, Clerk+JWT dual-auth | | Frankie | frankie-frontend-insights.md | ~15KB | Named CSS class architecture, brand.ts pattern, GSAP primary | | Shakespeare | shakespeare-content-insights.md | ~12KB | 6-tier blog architecture, PAA clustering, Ranking Reels pipeline | | Ghost | ghost-local-seo-insights.md | ~20KB | 95/4/1 trust fortress, 25/46 page targets, full schema stack | | Tommy | tommy-techseo-insights.md | 22KB | 22-check audit framework, 300+ patents, CTRify 7-phase | | Knox | knox-infra-security-insights.md | ~15KB | VisionClaw key exposure flagged, Cloudflare tunnel, Docker patterns | | Willie | willie-wordpress-insights.md | ~12KB | Hosting infrastructure, Elementor kits, WP-to-Next.js migration | | Gino | gino-ghl-insights.md | ~10KB | 29-snapshot audit, DFY pipeline, private API auth capture | | Picasso | picasso-design-insights.md | ~12KB | MDW Aesthetics brand system, Banana Squad pipeline, Creatify workflow | | Vox | vox-telephony-insights.md | ~18KB | 13 call flows, 740 LOC twiml-handler, $27 for 31K calls | ## How to Use Each file is a standalone reference. When an agent needs historical context for their domain, read their corresponding extract file before starting work. **Why:** These are the things the agents learned from mining months of conversation history. They contain decisions, patterns, anti-patterns, and preferences not derivable from code alone.

## 7. Anti-Patterns ### GMB / Profile Anti-Patterns - **Keyword stuffing the business name** — highest suspension risk factor. "Best Plumber Houston TX 24/7 Emergency" is a policy violation and a spam signal. Business name must match the legal or commonly known name exactly. - **Inconsistent hours** — GMB hours that don't match website hours confuse both users and Google's entity consolidation. Update both simultaneously. - **Identical descriptions across locations** — Google detects duplicate content in GMB descriptions. Each location profile needs a unique description. Even minor paraphrasing is insufficient — write fresh for each location. - **Same photos on all locations** — photos are an entity signal. Using identical photo sets across 10 location profiles dilutes their uniqueness signal. Each location should have location-specific photos. - **Leaving Q&A unmonitored** — competitors and spammers can post misleading Q&As on your GMB. Seed it with your own content and monitor weekly. - **Letting posts go dormant** — a GMB with no post activity for 90+ days signals an inactive business. Even a single weekly post prevents this. ### Citation / NAP Anti-Patterns - **Abbreviation inconsistency** — "St" vs "Street" vs "St." creates three different NAP strings. Google's entity matching is string-sensitive. Pick one format and apply it everywhere. - **Toll-free number in citations** — 1-800 numbers are not local signals. Use a local-format number as the primary NAP phone even if the business also has a toll-free line. - **Directory listing without photo** — citation entries with no photo get lower confidence scores from data aggregators. Always upload the logo or a business photo. - **Letting aggregator submissions sit without verification** — Data Axle and Neustar Localeze allow competitors to "suggest edits." Check your aggregator listings every quarter. ### On-Page / Content Anti-Patterns - **Stock photos** — never. Unsplash, Pexels, Pixabay, or any free stock service is prohibited. Only: Google Gemini AI-generated, client-provided, or CSS gradient placeholders. - **Generic "click here" or "learn more" anchor text on service links** — wastes internal link equity. Use descriptive keyword-rich anchors: "flat rate movers in Sarasota" not "click here." - **Same testimonials across all location pages** — if a client left a review about their Sarasota move, use it only on the Sarasota pages. Using it everywhere dilutes uniqueness. - **City name only in H1** — "Botox in Miami" as a lone H1 is thin. H1 should include the primary differentiator: "Botox Injections in Brickell Miami — Luxury Pricing, Expert Injectors." - **Arbitrary word count floors** — blog post length must match the #1 ranking competitor +10-20%. Never set a blanket 1,500-word floor if the top-ranked post is 800 words. - **No breadcrumbs on subpages** — breadcrumbs are both a UX signal and a BreadcrumbList schema opportunity. Every subpage (services, areas, blog posts) gets breadcrumbs. - **Missing canonical tags** — city-service intersection pages can create thin content at scale. Canonical tags prevent duplication penalties when content is similar across multiple city variants. ### Schema Anti-Patterns - **Generic Organization type** — using `@type: Organization` when the business is a local service provider wastes specificity. Always use `LocalBusiness` as the base, then add the niche subtype. - **Mismatched schema and page content** — if the schema says `aggregateRating: 4.9 / 789 reviews` but the page shows different numbers, it is a consistency violation. Keep schema data in sync with visible page content. - **No @id on LocalBusiness** — without `@id`, each page's schema is a disconnected entity. The `@id` URI ties all schema across the site to one canonical business entity. - **FAQPage schema on pages with no visible FAQ section** — Google requires FAQ schema to match the visible on-page content. Hidden or missing FAQ sections with schema present = quality violation. ### Site Architecture Anti-Patterns - **Catch-all Vercel rewrite on static sites** — wildcard `rewrite` in vercel.json causes soft 404s on dynamically-named city-service pages. Use explicit routes or static generation. - **Sitemap not covering all pages** — if the sitemap generator only tracks pages it personally created and misses pages generated by other build scripts, Google won't crawl the full site. Scan the output directory, not the internal page registry. - **Two separate build scripts with different styling** — if main template pages and city-service subpages are generated by different scripts, they often have design discontinuities. Single build script eliminates this. ### Multi-Location Anti-Patterns - **Cross-location phone number contamination** — Location A's phone number appearing on Location B's citation or page. This confuses call attribution and dilutes NAP trust for both locations. - **Identical UTM parameters across locations** — if both Sarasota and Bradenton GMBs use the same UTM tags, GA4 cannot distinguish traffic sources. Unique `utm_content` per location is mandatory. - **Reusing the same tracking number across inactive/active GMBs** — Mike manages 220+ GMBs with 278 SignalWire numbers and 52 CallFire numbers. Tracking numbers must be 1:1 with GMB profiles. Recycling a number from a suspended GMB to an active one can carry negative history. ---

## 6. Local Link Building ### Trust Fortress Architecture (Greg Montoya, SEO Rockstars 2015) The trust fortress approach treats link equity like a security perimeter: - **95%** of PBN links point to Tier 1 citations (Yelp, YP, BBB, directories) - **4%** of links point to Tier 2 web 2.0 properties (WordPress.com, Blogger, Tumblr) - **1%** of total links point directly to the money site Rationale: pointing PBN links directly at the money site leaves a visible footprint. Layering through citations and web 2.0s dilutes the footprint while still passing equity. ### Weebly 301 PBN Strategy (Carolyn Holzman, SEO Rockstars 2016) - Build Weebly microsites as buffer layer between PBN and money site - 301 redirect Weebly sites to the money site or Tier 1 citations - Carolyn's case study: 22,000+ backlinks acquired via this method - Risk profile: Weebly free sites are fragile — Weebly can deactivate them. Use paid Weebly accounts or mirror on GitHub Pages as backup. - 2024+ note: this is a gray-to-black hat tactic. Apply carefully in competitive niches where organic link acquisition is slow. ### Social Signals and Co-Citation (Josh, SEO Rockstars 2016) - Co-citation: getting your business name and URL mentioned together (without a link) on authoritative pages still passes relevance signal. - Social signals: Facebook shares, Google+ citations (historically), Twitter embeds — drive crawl budget and act as co-citation. - For local SEO, co-citation in local news articles and neighborhood blogs carries more weight than social share volume. - Strategy: get mentioned (even without a link) on local chamber of commerce sites, local news outlets, neighborhood association pages. ### NoFollow Links as Trust Signals (Dori Friend, SEO Rockstars 2015) - As of 2015 (the year of this data), NoFollow links from high-authority domains were more valuable as trust signals than DoFollow links from low-authority domains. - The trust transfer via NoFollow is not PageRank — it is entity co-occurrence and topical relevance. - Practical application: a NoFollow mention from a city government site, hospital, or university carries more trust signal than a DoFollow link from a low-DA directory. ### Link Silo Architecture (Moon Hussain, SEO Rockstars 2020) - Site structure and link silos were tested across 50 sites: 75% of sites ranked for 200+ keywords within 4 months with zero backlinks. - The silo structure creates internal link equity concentration: hub page accumulates PageRank from all subpages, then redistributes it via contextual links. - For local sites: service hub links to all service pages; service pages link back to hub and cross-link to adjacent services. - Never link from a service page to a competing service page's primary keyword anchor — use branded or generic anchors for cross-links. ### Negative SEO Defense (Ted Kubaitis, SEO Rockstars 2014 / Josh, 2016) - Monitor backlink profile monthly (Ahrefs/Majestic). Sudden spike in spammy backlinks is a negative SEO attack. - Disavow file: submit a disavow file to Google Search Console within 30 days of attack detection. Do not wait. - Velocity attacks: artificial link spikes timed with Googlebot crawl cycles (Ted Kubaitis 2014) — the penalty hits during the next quality signal refresh. - Fictitious URL attacks: spammers create 404 URLs on your domain and point spam links at those URLs. Monitor Search Console for crawl errors spiking on URLs you didn't create. ---

## 3. Geogrid & Maps ### Geogrid Fundamentals - Tool: LocalViking is the primary geogrid rank tracker in use. - Grid setup: center point on the business address. Grid size depends on service area — 7x7 for tight urban, 13x13 for suburban, 21x21 for wide rural service area. - Reporting cadence: baseline on day 1, then monthly delta report. Tie every rank movement to specific optimization action for attribution. - Proximity fact: you cannot change proximity. Google uses the searcher's location relative to the business. Accept this and focus on relevance + prominence signals to compensate. - Rank tie-breaking: when two businesses have equal proximity, relevance (categories, keywords, content) and prominence (reviews, links, citations) determine which wins the tie. ### Maps Pack Local Ranking Signals (Hierarchy) Based on pattern from conference SOPs and client data: 1. Google Business Profile completeness and category accuracy 2. Review velocity, sentiment, and keyword density in review text 3. On-page relevance of linked website (especially service pages) 4. Citation volume and NAP consistency 5. Backlinks to GMB-linked website 6. GMB post activity (recency signal) 7. Photo count and engagement 8. Proximity to searcher (cannot be changed) 9. Behavioral signals: clicks, direction requests, calls from GMB (CTR signals) ### CTR Manipulation Signals (Brent Bowser / CTR Geeks) - Brent Bowser built CTR Geeks as a tool for generating traffic signals to GMB listings and local pages. - Signal types: direct searches for the brand name + service, maps views, direction requests, calls from GMB, website clicks from GMB. - CTR Geeks course has 4 documented testimonials (Graham, Daryl Osborne, Bradley, James Slattery). - Pattern: simulated behavioral signals (views, clicks, calls) treated by Google's ranking algorithm as engagement indicators, which boost map pack position. - This is a gray-hat tactic — use for competitive markets where organic signals alone are insufficient. ### Maps Spam Fighting (Stephen Floyd SOP, 2020) Full SOP extracted from 2020 SEORockstars conference: 1. Identify spam: business name keyword stuffing (e.g., "Best Plumber Houston - 24hr Emergency"), fake address (residential or UPS Store), duplicate listings, fake reviews. 2. Reporting method: use "Suggest an edit" on the GMB listing to flag inaccurate business names. Report fake addresses directly via the "Report a problem" link on the map pin. 3. Mass reporting: if a competitor has 5+ spam indicators, report each issue separately — Google processes individual reports, not aggregate complaints. 4. Evidence package: screenshots of the listing with spam indicators highlighted, street view showing no business signage, reverse address lookup showing it's a residential or mail drop. 5. Success rate: Stephen Floyd reported 60-70% success rate on properly documented spam reports. 6. Timeline: Google typically reviews reports within 2-4 weeks. Follow up at day 30 if no action. 7. Maps spam tracker: maintain a spreadsheet of all spam reports submitted — date, listing name, CID, issue type, status. ### KML Files for Service Area - KML files define a custom service area polygon for GMB profiles that serve customers at their location (vs storefront). - Use Google My Maps to draw the service area polygon, export as KML, then upload to the GMB service area settings. - The KML polygon should match the service area pages built on the website — geographic consistency between GMB and on-page signals reinforces the relevance signal. - For moving companies: draw KML to cover origin + destination zones, not just a radius from the business address. ---

## 2. Citation & NAP ### Primary Citation Platforms (13 Tier 1) Priority order for new client builds (from Flat Fee Movers deployment checklist): 1. Google Business Profile (GMB) — if not done, start here 2. Yelp — claim/create, add full description, photos, categories, hours 3. BBB (Better Business Bureau) — paid accreditation if budget allows; drives trust signals 4. Facebook Business Page — complete all fields, add services section 5. LinkedIn Company Page — especially for B2B or professional services 6. Thumbtack — service area + categories + intro message 7. YellowPages — two separate entries possible (YP.com + superpages.com parent) 8. Superpages — separate submission from YellowPages despite shared ownership 9. Angi (formerly Angie's List) — moving, home services, trades especially 10. HomeAdvisor — service area, categories, request quotes enabled 11. Nextdoor Business — neighborhood-level visibility, very high trust in home services 12. Houzz — relevant for home services, remodeling, interior 13. MapQuest — still indexed by Bing/Apple Maps data pipelines ### Data Aggregators (4 mandatory) These seed dozens of downstream directories automatically. Submit here first, then Tier 1 manually. 1. Data Axle (formerly Infogroup) — dataaxle.com/business-listing 2. Neustar Localeze — neustar.biz/localeze 3. Foursquare — foursquare.com/business 4. Yelp — also functions as aggregator for Apple Maps data ### NAP Consistency Rules (5 rules) 1. Exact match to GMB — copy-paste the exact business name, address, and phone from the GMB profile. No abbreviations, no variations. 2. No cross-location mixing — each location gets its own NAP block. Never use Location A's phone number on Location B's citation. 3. USPS format for address — use standard USPS postal abbreviations (St not Street, Ave not Avenue). Verify with USPS address lookup tool. 4. Local phone number only — no toll-free numbers as the primary NAP phone. Call tracking numbers are acceptable if they are local-format and consistently applied. 5. Quarterly NAP audits — run a citation audit every 90 days minimum. Use BrightLocal or Whitespark. Fix any drift immediately. ### Canonical NAP Block Format ``` Business Name: [Exact legal name or DBA matching GMB] Address: [Street Number] [Street Name Abbrev], [City], [State] [ZIP] Phone: ([area code]) [3-digit]-[4-digit] Website: https://[domain.com]/?utm_source=citation&utm_medium=[directory] ``` ### Duplicate Detection - Same business name with different CID URLs (goo.gl vs google.com/maps?cid=) signals a duplicate or split listing. Merge or suppress the legacy listing. - Common duplication source: old GMBs created pre-2018 with short URLs. Audit these in the GMB database by checking for name collisions with different CID values. - Google's duplicate detection is imperfect. If two listings exist for the same address, the one with more engagement signals (reviews, posts, photos) typically wins — but both should be manually merged. ---

## 1. GMB Optimization ### Profile Completeness - Business description: max 750 characters. Formula — lead with primary keyword + city in sentence 1, then 2-3 service differentiators, close with call-to-action sentence. Never keyword-stuff sentence 1. - Q&A seeding: pre-populate 5-10 high-volume PAA questions as Q&As on the GMB profile. Own the Q&A section before competitors or spammers do. Include target keyword naturally in each answer. - Photos: minimum 10 at launch. Always include interior, exterior, team, and work-in-progress shots. Zero stock. Geotagged JPEGs only (embed GPS coordinates in EXIF before upload). - Attributes: fill every applicable attribute — wheelchair accessible, women-owned, veteran-owned, online estimates, etc. Attributes drive filter visibility in local pack. - Categories: primary category is the single most important ranking lever. Secondary categories add topical relevance but do not dilute the primary. Research competitors' primary categories before selecting. ### GMB Post Strategy - Post type: EVENT or OFFER type consistently outperforms WHAT'S NEW for engagement signals. - Duration: set end date 1 year out for evergreen offers. Never let posts expire without replacement. - Title: keyword-first, not brand-first. "Brooklyn Flat-Rate Movers — Spring Booking Now Open" not "Buy The Hour Movers Announces..." - Body: 150-300 words. Weave in 2-3 secondary keywords naturally. Include a neighborhood reference. - CTA buttons: "Book now" for service/transaction posts, "Call now" for consultation/emergency posts, "Learn more" for informational/blog posts. - Multi-location brands: write separate post versions per location. Never post identical text across both GMB profiles. Location-specific phone numbers, neighborhood references, and seasonal context must differ. - UTM tagging: append UTM parameters to all GMB post links. Pattern: `utm_source=gmb&utm_medium=post&utm_campaign={post-topic}&utm_content={location-slug}`. Enables GA4 attribution by location. - Velocity: at least 1 post per week. Dormant profiles lose ground in map pack. Consistent posting = activity signal. ### Review Strategy (from Mike's $100K testing data) - Reviews with keywords + photos outperform links as local pack signals. - Review sentiment is the strongest local ranking signal — outweighs entity authority. - Velocity matters more than total count. 5 reviews/month beats 50 reviews once with nothing since. - Review response: respond to every review within 48 hours. Include the business name + city + service in the response naturally ("Thanks for trusting Flat Fee Movers Sarasota with your move..."). - Prompting: give customers a direct link to the GMB review form. Never use a redirect or landing page in between — friction kills conversion. - Keyword-rich review prompting: ask customers to mention the specific service and city in their review. Example prompt: "If you'd like to help others find us, mentioning the specific service you used (e.g., 'local moving in Sarasota') would be especially helpful." ### UTM + CID Tracking - Every GMB profile should have a UTM-tagged website URL. Pattern: `?utm_source=gmb&utm_medium=organic&utm_campaign=local-pack&utm_content={location-slug}` - CID URL format: use `https://www.google.com/maps?cid={CID_NUMBER}` — this is the canonical format. Avoid goo.gl short URLs (legacy, unreliable for tracking). - CID extraction: CIDs are retrievable from the GMB URL when you click "Share" on a listing, or via Google Places API. - Multi-location database (from Mike's 540-record GMB DB): track CID URL, website URL, phone, GMB status (active/inactive/suspended), last verified date, and call tracking number per profile. ### Multilogin Browser Profiles - Assign a dedicated Multilogin browser profile per brand. Store the profile ID in the brand's database record. - One-click launch from domain portfolio dashboard — prevents accidental cross-account contamination. - Never manage multiple client GMBs from the same browser session or IP. ### Suspension Recovery - GBP Recovery Coach approach: 160+ item interactive checklist, 7 chapters (pre-suspension audit, during-suspension triage, evidence gathering, appeal writing, reinstatement, post-reinstatement hardening, monitoring). - Evidence kit: business license, utility bill showing business address, lease agreement or deed, photos of physical signage, supplier invoices with business name + address. Date-match all documents. - Appeal letter formula: state the business name and category, provide timeline of operation, attach evidence, cite the specific policy you believe you complied with, and offer to provide additional documentation. Never admit guilt or speculation. - Reinstatement rate improves significantly when the appeal is submitted within the first 30 days of suspension. - Post-reinstatement: immediately update all citation NAP to match reinstated GMB exactly. One mismatch can trigger re-review. - RAG AI coach (from PRD): OpenAI-powered coach that answers questions contextually from the 7-chapter handbook and checklist state. ---

## 2. Engineering Patterns ### 2.1 /prime Command -- Context Acquisition Before Work **Pattern:** Every project session starts with `/prime` which reads project structure, CLAUDE.md, package.json, brand assets, git context, and outputs a Project Context Summary card. **Why:** Agents without context produce garbage. The prime command is the single most important step -- it loads everything the agent needs before touching code. **How it works:** Reads key files in parallel (git ls-files, README, package.json, brand guide, env examples), identifies tech stack and project type, outputs a formatted reference card. **Status:** Mature. Evolved from basic session recap to full "Context Acquisition + Reasoning Engine." ### 2.2 Quality Gates System **Pattern:** Phase-gated development where each phase has explicit checklist criteria that must pass before advancing. **Why:** Prevents cascading failures. Found in micro-saas-orchestrator and aka-orchestrator but originally missing from the universal Oliver template. **Example gates:** Phase 1 (PRD reviewed, architecture documented, schema designed, API endpoints defined) -> Phase 2 (project bootstrapped, DB tables created, RLS in place, dev environment working). **Status:** Identified as critical missing feature in Jan 2026 template audit. Being retrofitted. ### 2.3 HANDOFF.json for Session Continuity **Pattern:** Structured JSON file capturing exact project state for cross-session handoffs: completed tasks, remaining tasks, blockers, human actions pending, decisions made, uncommitted files, next action. **Why:** Claude Code sessions are ephemeral. Without structured handoffs, context is lost between sessions. HANDOFF.json is machine-readable, enabling automated resume workflows. **Evidence:** ClaudeClaw project uses this extensively. GSD framework's `/gsd:resume-work` command reads HANDOFF.json + .continue-here.md to reconstruct state. **Status:** Active pattern. Complemented by .continue-here.md (human-readable summary) and STATE.md (broader project state). ### 2.4 3-Agent Auto-Scaffold Pipeline **Pattern:** credential-provider -> infrastructure-initializer -> new-project-orchestrator. Three agents that handle new project creation end-to-end. **Why:** Manual project setup (copying .env files, creating Supabase projects, configuring Clerk) is repetitive and error-prone. The pipeline reads from MASTER_API_KEYS_COLLECTION.env and creates project-specific configs automatically. **Evidence:** Built in Feb 2026 session. Oliver says "create a micro-SaaS with Supabase + Clerk + Stripe" and the pipeline handles everything. **Status:** Built. Three agent files + command reference + bootstrapping guide created. ### 2.5 Multi-Access Pattern for Tools **Pattern:** Each tool/skill can be accessed as Agent (persistent, context-retaining), Skill (one-shot command), or Dashboard Tool (form-based UI) depending on the use case. **Why:** Different interaction patterns serve different needs. A blog monitor needs persistent context (agent mode), but a quick RSS check is a one-shot command (skill mode), and feed configuration benefits from a form UI (dashboard mode). **Classification matrix:** Each of 9 core tools was rated for primary/secondary/tertiary access patterns based on: complexity of input, need for visual output, context retention requirements, real-time interaction needs. **Status:** Designed in Feb 2026. Classification matrix created. ### 2.6 Agent-to-Skill Conversion Pattern **Pattern:** Community agent templates (full personality + JSON protocol + inter-agent references) are converted to skill files by: stripping agent personality/greeting instructions, removing JSON communication protocol payloads, removing inter-agent references, adding proper frontmatter (name, description, category, tags), adding "When to Use" / "When NOT to Use" routing sections, preserving all technical content verbatim. **Why:** Community templates are designed as standalone agents. Mike's system needs them as reference skills that other agents can invoke. The conversion preserves knowledge while fitting the ecosystem's access pattern. **Evidence:** Mar 2026 session converting python-pro and golang-pro agent templates to skills. Systematic process applied across multiple conversions. **Status:** Active pattern. Applied to 60+ community templates. ### 2.7 GSD Framework (Get Shit Done) **Pattern:** 12-agent framework with specialized roles: codebase-mapper, debugger, executor, integration-checker, nyquist-auditor, phase-researcher, plan-checker, planner, project-researcher, research-synthesizer, roadmapper, verifier. Each has a corresponding workflow skill. **Why:** Breaks down complex project work into discrete, specialized phases. Each agent has a narrow scope with clear inputs/outputs. **Key insight:** Most gsd-* skill names listed in agent frontmatter do NOT exist as actual skill directories -- they are orphaned skill references. The workflow definitions need to be built. **Status:** Framework defined, partially implemented. Workflow skills are the gap. ### 2.8 Blueprint Engine Pattern **Pattern:** Structured workflow definitions (JSON blueprints) that match incoming tasks to predefined step sequences. Steps can be: agent (spawn a specialist), code (execute a script), or gate (require human approval). **Why:** Moves from free-form "Oliver figures it out" to structured, repeatable execution paths. Blueprint matching enables consistent quality for common task types (build client site, run SEO audit, do keyword research). **How it works:** Task feeder polls for new tasks -> matches against blueprint library -> if match, runs structured blueprint with step executor (retry logic, expertise writeback) -> if no match, falls back to free-form Oliver. **Status:** Deployed. Three starter blueprints (build-client-site, seo-audit, keyword-research). Runner on VPS. ### 2.9 Smart Chunking by File Type for RAG **Pattern:** Different file types require different chunking strategies for effective vector search: - Python/JS: chunk by function/class boundaries - Markdown/transcripts: chunk by headers/sections/paragraphs - YAML configs: keep whole or chunk by top-level keys - JSON: chunk by top-level objects **Why:** Naive chunking (fixed character count) produces garbage fragments. A function split in half is useless. A YAML file split mid-key breaks context. Smart chunking preserves semantic units. **Status:** Designed, partially implemented as part of RAG pipeline. ---

## 1. Architecture Decisions ### 1.1 Orchestra Model for Agent Orchestration **Decision:** Adopt a composer/conductor/section-leader/musician hierarchy instead of flat agent dispatch. **Why:** Flat dispatch (Oliver routes directly to 50+ specialists) created a routing nightmare -- too many agents to reason about, too many handoff errors. The orchestra model separates concerns: Oliver writes the brief score (what, who, constraints), Carlos conducts (sequencing, dependencies, parallel safety), leads plan their own domain, specialists execute. **Evidence:** Multiple sessions show Mike correcting Oliver for writing overly detailed workstreams ("the drummer writes the drumming part"). Oliver's scope was explicitly narrowed to brief scores: objective, constraints, which leads, done-when. **Status:** Active. Enforced in oliver-soul.md. ### 1.2 Subagents Over Teams as Default **Decision:** Use Claude Code Agent tool (subagents) instead of Teams for all work unless Mike explicitly says "team." **Why:** Teams are ~5x more expensive (persistent parallel sessions billing continuously). Subagents spawn, do work, die. Same memory capability via backup stack. Teams only justified for long-horizon parallel campaigns or peer-to-peer messaging scenarios. **Evidence:** Apr 11 2026 revision in oliver-soul.md. Prior sessions showed unnecessary Team spawning burning tokens on simple tasks. **Status:** Active. Hard rule. ### 1.3 Thin Template + Centralized Library **Decision:** Oliver template is a minimal bootstrap shell; all agents, skills, scripts, standards live in a centralized "Library of Claude Codeia." **Why:** Duplicating 50+ agents and 693+ skills across every project creates maintenance hell. Central library means updates propagate instantly, no duplication, single source of truth. Template just has config pointers + hooks + session memory. **Evidence:** Feb 2026 session where Mike corrected the approach: "the whole point of my template is [not to embed everything]." Template reads library path from config, dynamically loads at runtime. **Status:** Architectural principle. The library lives at `D:\ClaudeDev\0_GITHUB\Tools\Library of Claude Codeia`. ### 1.4 Pinecone + OpenAI Embeddings for Drive-Wide RAG **Decision:** Selected Pinecone (serverless) + OpenAI text-embedding-3-large over Supabase pgvector, ChromaDB, Weaviate, or Qdrant for full D: drive RAG. **Why:** Zero infra management (enough servers to babysit already). Namespace per folder category. Metadata filtering for file-type-specific search. Scales to millions without thinking. Supabase pgvector is fine for small RAG but not purpose-built for "dump my entire drive" scale. ChromaDB is prototyping-grade. Weaviate and Qdrant are strong alternatives but add infra burden. **Evidence:** Mar 2026 Agent-Soul-System session with full comparison matrix. Estimated 2-hour build time for ingestion pipeline. **Status:** Decision made, pipeline partially built. Key insight: smart chunking by file type is the hard part -- code chunks by function boundaries, markdown by headers, YAML keeps whole. ### 1.5 Convex for Mission Control Backend **Decision:** Use Convex (real-time backend) for the mission control / blueprint engine. **Why:** Provides real-time reactivity for dashboard state, structured schema with indexes, HTTP routes with CORS, and serverless execution. Blueprint engine stores structured workflow definitions (blueprints) that match incoming tasks to predefined step sequences. **Evidence:** Mar 2026 session deploying blueprint engine -- Convex schema + functions for blueprints + blueprintRuns tables, 6 HTTP endpoints, Python runner on VPS. **Status:** Deployed. Blueprint-aware task feeder polls Convex, matches incoming tasks to blueprints, runs structured execution or falls back to free-form Oliver. ### 1.6 One Discord Bot Per Agent (Not Channel Routing) **Decision:** Each of Mike's 19 core agents gets its own Discord bot identity, not a single bot routing messages by channel. **Why:** Mike wants each agent to have its own identity, DMs, and presence in Discord. Individual presence matters more than implementation simplicity. **Evidence:** ClaudeClaw setup session, Apr 2026. 18/19 bots brought online; remaining needed Discord Developer Portal manual fixes (Message Content Intent). **Status:** Implemented. Blocked on manual Discord portal configuration for 3 bots. ### 1.7 Tailscale Funnel Over Cloudflare Tunnel **Decision:** Use Tailscale Funnel for permanent dashboard URLs instead of Cloudflare tunnel. **Why:** merlinoai.com is on Namecheap not Cloudflare. Funnel works immediately with no DNS changes. Zero configuration overhead. **Evidence:** ClaudeClaw session, explicit decision logged in HANDOFF.json. **Status:** Active. Dashboard accessible via Tailscale Funnel URL. ### 1.8 Next.js + Supabase + Clerk as Default Stack **Decision:** Default application stack: Next.js (App Router) + Tailwind CSS + ShadCN + Supabase (Postgres + Auth + Storage) + Clerk for auth when needed. **Why:** Standardization across all client and internal projects. Deploys to Vercel. Mike knows the stack. Every agent knows the stack. Reduces context-switching costs. **Evidence:** Multiple project scaffolding sessions (ADHD todo, SEO SaaS suite, GMB tools). Consistent across CLAUDE.md global instructions. **Status:** Active default. Override only when Mike specifies. ---

# Dan Architecture Insights > Extracted from chat backup archive: Agent-Soul-System, Coding-Projects, Claude-Tools, Oliver-Orchestrator, ClaudeClaw > Source: ~300+ chat transcripts spanning 2025-11 through 2026-05 > Extracted: 2026-05-09 ---

## 4. Anti-Patterns ### 4.1 Oliver Doing Work Instead of Delegating **What happened:** Oliver would write code, create content, run deploys -- doing specialist work instead of orchestrating. **Why it's wrong:** Violates the orchestra model. Oliver should write the brief score, not play every instrument. When Oliver does work, it burns context on tasks a specialist agent could handle better. **Fix:** Hard rule: "If you catch yourself about to DO work instead of DISPATCHING -- STOP." Enforced via soul file. ### 4.2 Embedding Everything in the Template **What happened:** Early Oliver template versions embedded all 48+ agent definitions, all skill references, all standards directly in the template. **Why it's wrong:** Creates maintenance nightmare. Every project has its own stale copy. Updates don't propagate. Bloats context for every session. **Fix:** Thin template + centralized library. Template is bootstrap only; everything else loaded from library at runtime. ### 4.3 Oliver Writing Domain-Specific Workstreams **What happened:** Oliver would write detailed implementation plans for frontend, SEO, content -- domains owned by leads. **Why it's wrong:** "The drummer writes the drumming part." Oliver doesn't have domain expertise. Detailed plans from a generalist are worse than brief scores that let specialists plan their own work. **Fix:** Oliver's plans are BRIEF: objective, constraints, which leads, done-when. No workstreams, no implementation steps, no technical plans. ### 4.4 Supabase pgvector at Scale **What happened:** Considered using existing Supabase pgvector for full D: drive RAG (thousands of files). **Why it's wrong:** pgvector is NOT purpose-built for vector search at scale. SQL-based filtering is slower than native vector DBs. No built-in chunking or hybrid search. Scaling past 500K vectors gets painful. **Fix:** Moved to Pinecone (purpose-built, serverless, scales to millions). Supabase pgvector is fine for small RAG but not for "dump my entire drive." ### 4.5 Orphaned Skill References **What happened:** 59% of skills (409 of 693) have zero agent references. GSD framework agents reference workflow skills that don't exist as actual skill directories. **Why it's wrong:** Dead references waste scanning time, create false expectations, and make the skill ecosystem harder to navigate. **Fix:** Audit and either: (a) wire orphaned skills to appropriate agents, or (b) archive/delete truly unused skills. Build missing GSD workflow skills. ### 4.6 Missing Quality Gates in Universal Template **What happened:** Domain-specific orchestrators (micro-saas, aka, kanban, sugar) had quality gates, status templates, decision trees, execution checklists -- but the universal Oliver template had none of these. **Why it's wrong:** The universal template was "significantly more basic" than specialized orchestrators. Projects using the universal template had no phase gates, no structured status reporting, no complexity assessment framework. **Fix:** Retrofit all missing patterns from specialized orchestrators into the universal template: quality gates, status reporting, task complexity assessment, orchestration patterns library, decision trees, pre/during/post checklists, metrics tracking. ### 4.7 Teams for Everything **What happened:** Before the Apr 2026 revision, Teams were used for tasks that subagents could handle. **Why it's wrong:** Teams are ~5x more expensive. Persistent parallel sessions bill continuously while alive. Inter-teammate SendMessage reloads burn tokens. For most tasks, subagents (spawn, work, die) achieve identical results at a fraction of the cost. **Fix:** Subagents are the default. Teams only when Mike explicitly says "team" or for long-horizon parallel campaigns requiring live peer-to-peer messaging. ### 4.8 $CLAUDE_PROJECT_DIR on Windows **What happened:** Hooks using `$CLAUDE_PROJECT_DIR` failed on Windows because the environment variable doesn't expand in the hook's execution context. **Why it's wrong:** Blocks file operations silently. The hook error is "non-blocking" in theory but actually prevented writes from going through. **Fix:** Use Bash tool (has explicit permissions) or fix the env variable expansion for Windows. This is a known Claude Code platform issue. ---

## 3. Agentic Infrastructure ### 3.1 Agent Ecosystem Scale **Inventory (as of Mar 2026):** - 202 agent definition files in ~/.claude/agents/ - 693 skill directories in ~/.claude/skills/ - ~160 agents reference skills; 28 agents spawn subagents - 284 skills referenced by agents (41%); 409 orphaned skills (59%) **Team structure:** - Core Team (Oliver's direct staff): 8 agents - Revenue (Sally's team): 5 agents - Marketing (Cole's team): 15 agents - SEO (Ghost's team): 18 agents - Dev (Merlin's team): 7 agents - GSD Framework: 12 agents - Builder/Specialist: 60+ agents **Key finding:** 59% skill orphan rate. More than half of all skills are not referenced by any agent. This is either dead weight or untapped capability. ### 3.2 ClaudeClaw -- Claude Code on Your Phone **What it is:** Not a chatbot wrapper. Spawns the actual `claude` CLI on Mac/Linux and pipes results back to Telegram. Everything that works in terminal -- skills, tools, context -- works from phone. **Architecture:** TypeScript core (bot.ts, agent.ts, config.ts, dashboard.ts, discord.ts, scheduler.ts, memory.ts, voice.ts). SQLite for state. War room (Python) for multi-agent voice interface. Dashboard on port 3141 via Tailscale Funnel. Discord integration with per-agent bots. **Key components:** Telegram bot, Discord multi-bot, dashboard HTML, scheduler (cron-based), vector memory, Obsidian integration, Slack integration, WhatsApp bridge, voice pipeline. **Status:** Active development. 7/10 setup tasks complete. Dashboard live, 18/19 Discord bots online. ### 3.3 Mission Control + Task Feeder Architecture **Pattern:** Convex backend (real-time dashboard) -> Task Feeder on VPS polls for new tasks -> Blueprint matching -> Blueprint Runner executes structured workflows OR free-form Oliver handles unmatched tasks -> Results written back to Convex. **Components:** - Convex schema: blueprints table (workflow definitions) + blueprintRuns table (execution history) - task-feeder-v3.py: Blueprint-aware poller - blueprint-runner.py: Step executor with agent/code/gate types, retry logic, expertise writeback - push-blueprints.py: Seeds blueprint definitions from JSON **Status:** Deployed on Hostinger VPS. ### 3.4 Memory Architecture Layers **Hierarchy (most to least frequently accessed):** 1. Always-loaded: MEMORY.md (agent memory index), soul files 2. Structured files: RECENT_CONTEXT.md, daily logs 3. Expertise YAMLs: smart-routed by CWD 4. Vector search: on-demand via agent-memory.py 5. Hindsight bank: cross-session recall by tags **Key principle:** Project state lives in the project (CLAUDE.md, .planning/STATE.md). Global = identity only. Never duplicate project state in global memory. ### 3.5 Hooks Infrastructure **Three hook types observed:** - context-guardian.js: Monitors context window usage, prevents overflow - damage-control.js: Prevents destructive operations - provenance-logger.py: JSONL session logging for audit trail **Problem found:** `$CLAUDE_PROJECT_DIR` environment variable doesn't expand properly on Windows, causing hooks to block file operations. Workaround: use Bash tool which has explicit permissions. ### 3.6 FastMCP for Custom MCP Servers **Decision:** FastMCP (Python framework) is the tool for building custom MCP servers. **Why:** Handles all protocol plumbing. Decorate Python functions to create MCP tools. Powers ~70% of MCP servers. 1M+ downloads/day. Versions: v1.0 merged into official SDK, v2.0 standalone with client support, v3.0+ current. **Status:** Available. Used for Context7 browser tools, Supabase, and other MCP integrations. ---

## Workflow Notes ### Image Generation Pipeline Order (Site Builds) 1. Phase 0: Pull existing site assets + CDN images to `public/images/site/` 2. Phase 4-5 (parallel with content): Generate Gemini site images 3. Phase 6: Map CDN URLs → local paths, verify alt text, zero stock photos 4. Generate before/after pairs: `scripts/generate-before-after.mjs` 5. Generate infographics: `scripts/nano-banana.mjs` 6. Generate Creatify video (optional, one per site, flagship service page) 7. Wire all media into pages before final build ### Banana Squad Usage Pattern - Invoke when: client wants high-quality hero image, OG image, or brand visual — not routine site images - Solo mode works identically to team mode — same quality, no team setup needed - After generation: Critic ranks all 5 variants → user picks → iterate on winner - Shutdown immediately after presenting results (token cost) ### Rate Limit Handling - Gemini 429: wait 60 seconds, retry - Never parallelize more than ~3 concurrent Gemini image requests - Space batch requests with small delay between calls

## Template & Mockup Patterns ### Website Starter Template - Location: `D:\Codeland2026\Templates\website-starter\` - Stack: Next.js App Router, TypeScript, Tailwind CSS 4 - Contains: globals.css base system, layout.tsx, brand.ts shell, FloatingCTA component - All sites copy this template, then customize brand.ts and globals.css ### Mission Control Dashboard Template - Full agency dashboard: agent cards, kanban, activity feed, revenue tracker, lead pipeline - Component styling: color-coded status dots (emerald = active, pulsing), scanline CSS header effect - MetricsBar: 5 metrics with emojis + icons, `tabular-nums` font for number alignment - Glass morphism on header: `backdrop-filter: blur(16px)` - Status pulse animations: agent active states use pulsing color-coded dots ### Self-Contained HTML Gallery Pattern - For design reviews: single HTML file with `@import` for fonts, all CSS inline - 8 unique font families to keep mockup options visually distinct - Used for client design approval flow before building in Next.js ### BeforeAfterSlider Component - React component for drag-to-compare before/after images - Service page placement: between Description section and What's Included section - `beforeAfterMap` TypeScript Record maps service slug → `{ slug: string; title: string }` - Image files: `public/images/before-after/{slug}-before.png` and `-after.png` ### infographicMap Pattern (Blog Posts) ```typescript const infographicMap: Record<string, string> = { 'blog-slug': 'infographic-filename', // ... }; // Render between article content and FAQ: // <Image src={`/images/infographics/${infographicMap[slug]}.png`} width={1080} height={1350} /> ``` ---

## Docker & Containers ### Clawdbot Docker Stack (microservices pattern) Services orchestrated via docker-compose: - Gateway (Port 8000): API router to all model services - Kimi (Port 3001): Moonshot K2.5 - Gemini (Port 3002): Google Gemini 1.5 Pro - Codex (Port 3003): OpenAI GPT-4 Turbo - Minimax (Port 3004): Minimax - Prometheus (Port 9090): Metrics Key patterns from this stack: - Docker network isolation: `clawdbot-network` with `driver: bridge` - Health checks implemented on all services - Service profiles for selective startup - CRITICAL FLAW: API keys hardcoded in docker-compose.yml (MOONSHOT_API_KEY, GOOGLE_GEMINI_API_KEY, etc.) — should use `env_file:` directive with a `.env` file ### Claude-Docker Dockerfile Patterns Non-root user with matching host UID/GID: ```dockerfile RUN adduser --uid ${HOST_UID} --gid ${HOST_GID} --disabled-password claude-user COPY .env /app/.env COPY .claude.json /tmp/.claude.json RUN cp /tmp/.claude.json /home/claude-user/.claude.json && \ chown -R claude-user /home/claude-user USER claude-user ``` Issue: `.env` baked into image at build time — acceptable for throwaway containers, not for persistent services. ### Mission Control (abhi1693/openclaw-mission-control) Enterprise control plane stack: FastAPI + Postgres + Redis + RQ workers + Next.js - Auth: Local bearer token or Clerk JWT - Deployment: Docker Compose on VPS - Connects to OpenClaw gateway via WebSocket - Provides approval workflows, audit logs, org/board/task management - vs. custom Convex template (push model, Vercel): the GitHub repo is a pull-based enterprise ops layer ### Container Security Checklist - Non-root user with specific UID/GID - No secrets in ENV vars or build layers (use `env_file:` + `.env`) - Read-only root filesystem where possible - Capabilities dropped: `--cap-drop ALL --cap-add NET_BIND_SERVICE` - Health checks on all services - Disable inter-container communication where not needed: `icc: false` in daemon.json - daemon.json hardening: `no-new-privileges`, `userns-remap`, log rotation, `live-restore` ---