plan · 2026-04-13

Who Is Mark Barnett — Research Plan

Status: Plan only. Deep research is ON HOLD until Mark provides personal info + verification tokens. Tasks #36 and #56.

Author: G16 Claude (Opus 4.6, 1M context)

Date: 2026-04-13

Audience: Future Claude sessions that need to load Mark as a person, not a profile.

0. What I already found without any help from you

Quick public sweep done before writing this plan, so you see real progress over your morning coffee.

GitHub MrB-Ed — public profile resolves. Display name "Mark Barnett," bio "Problem Solver," joined 2025-03-09, 2 public repos, 1 public gist, 0 followers, 0 following, no location/email/blog set.

MrB-Ed/The_Project — "The original personal AI memory and analytics project — predecessor to TP3 Neural Stack (TP2 era and earlier)." Python-heavy (72%), 15 commits, 10 PRs, last push 2026-04-10. Contains AGENTS.md, PROJECT_STATE.md, HANDOVER_CONTEXT.md, NETWORKING.md. This repo is publicly readable and already exposes the architecture and naming of the digital-twin work.
MrB-Ed/Soccer-MultiCam-Core — archived, "professional multi-camera soccer analytics," 7 KB, two commits in Feb 2026.
Public gist: "Proposal for Mark Barnett's Secure Data Management System" authored by "Antigravity (AI Assistant)," 2026-04-10. Three architectures (Vaultwarden+Bitwarden, Doppler/Infisical, AI Safe-Drop). Mentions G16, TP2, ChromaDB, .env migration, Ollama. This is currently public and names Mark, the G16 device, and the secrets-migration story.

thebarnetts.info — DNS resolves via Cloudflare but origin returns HTTP 522 (origin unreachable). Tunnel/origin is down. From the public internet today, no content is leaking — but the hostname itself confirms a Barnett-family domain on Cloudflare.

Saint Francis School / sfschools.net — no Mark Barnett matches in Google indexing for "Saint Francis" + Atlanta + history teacher. School is St. Francis Schools (Roswell/Milton, GA, founded 1976). No public faculty page hit on Mark's name. Public web does not currently expose his employment.

LinkedIn — no Mark Barnett + Saint Francis hit via Google site:linkedin.com. Either no profile, or profile is not indexed.

OMI community / breezybarnett16 — zero public hits. Username does not appear on indexed sites.

Bottom line of the recon: the only public surface today is the GitHub account and one gist. Both name "Mark Barnett" and disclose project architecture. Everything else (employment, family, OMI, Honest Answers, Legacy Soil) is currently not findable by a stranger Googling your name. That is the baseline — the deep research will go far past what a stranger can see.

1. Skills reconnaissance — is there a "person research" skill?

Anthropic official anthropics/skills — no person-research skill. Their public skills repo demonstrates art, music, MCP generation, web testing, enterprise workflows. Nothing for "profile a person."

Community ecosystems found:

alirezarezvani/claude-skills (232+ skills, multi-agent)
VoltAgent/awesome-agent-skills (1000+ skills aggregator)
travisvn/awesome-claude-skills, ComposioHQ/awesome-claude-skills, BehiSecc/awesome-claude-skills, glebis/claude-skills, wondelai/skills
sumba101/OSINT-AI-Agent — only direct hit. Skill osint-investigation wraps Holehe (email→accounts), Sherlock (username→400+ social sites), GHunt (Gmail intel). Hostile-OSINT framing, but the workflow (verify tools → email pass → username pass → JSON parse → markdown report) is exactly the shape we want, just inverted from adversarial to consensual.

Verdict: no off-the-shelf skill matches "friendly self-research on a consenting subject who hands you their accounts and tokens." Recommendation: build a custom skill named self-portrait-research, structured as:


~/.claude/skills/self-portrait-research/
  SKILL.md          # invocation + safety rules
  phases/
    00_recon.md     # public web, no auth
    01_arming.md    # accept token bundle from subject
    02_gmail.md     # Gmail MCP queries
    03_omi.md       # OMI transcript pulls
    04_github.md    # gh CLI + commit mining
    05_drive.md     # Drive doc inventory
    06_postgres.md  # TP3 Neural Stack queries
    07_synthesis.md # produce Who_Is.md with citations
  templates/
    who_is_template.md
    citation_format.md
  exclusions.md     # health, legal, family-private

The skill should refuse to run any phase past 00_recon without an explicit --consent-from <subject> flag pointing at a signed consent file. That makes it safe to publish later if you ever want to.

2. Data sources available with your consent (signal vs. noise)

Source	What I'd extract	Signal	Noise
Google Drive AI_Library	Project docs, research notes, daily compilations, Legacy Soil binders, Bidet AI specs	Your in-progress thinking, pricing decisions, voice	Old drafts, AI-generated filler, abandoned ideas
Gmail (MCP)	Senders you respond to, threads you start, school comms, vendor research, OMI/Anthropic receipts	Real relationships, real obligations, what you actually pay for	Newsletters, school admin spam, marketing
OMI transcripts	Recurring phrases, names you mention, who you talk to vs. about, emotional cadence, where you are when you think	Highest-fidelity portrait of how you actually talk	Ambient noise, half-thoughts, mishears
TP3 Postgres (Ollama 628,901 + Gemini 3,342 rows)	Topic clusters, time-of-day patterns, model-specific question styles	Quantitative shape of what you've been thinking about	Test runs, prompt experiments, duplicates
GitHub MrB-Ed	Commit messages, PR descriptions, README voice, repo naming patterns	How you describe technical work in your own words	AI-generated commit text
`~/.claude/projects/-home-g16/`	Session logs, what you ask Claude over and over, frustration markers	Your operating loop with AI tools	Tool-call noise
Social accounts (TBD by you)	Public posts, follow graphs, profile bios, photos	Who you choose to be in public	Algorithm-driven engagement

3. The Plan — phased

Phase 0 — Recon (DONE above). ~30 min. Tools: WebSearch, WebFetch, GitHub API, curl. Output: section above.

Phase 1 — Arming. Blocks on you. Provide: full DOB, cities lived in, social handles (X/IG/FB/Reddit/YouTube/Substack/etc.), aliases, OMI device ID, OAuth scopes you authorize, a consent file at ~/.claude/projects/-home-g16/memory/consent_self_portrait.md. Time on your side: 30 min of brain-dump. Output: arming_bundle.json.

Phase 2 — Gmail sweep. ~45 min. Tools: mcp__claude_ai_Gmail__gmail_search_messages, gmail_read_thread. Pull top 50 senders by thread count, top 20 senders you initiate to, label inventory, drafts (drafts reveal what you couldn't finish saying). Extract: people you actually correspond with, recurring topics, vendor/tool stack, subscriptions. Output: phase2_gmail.md.

Phase 3 — OMI transcript sweep. ~90 min. Tools: OMI API + TP3 Postgres queries against transcript tables. Strategy: (a) frequency of named people, (b) frequency of named projects, (c) sentiment shifts across days, (d) recurring phrases > 3 words ("digital twin architect," "patio life," etc.), (e) questions you ask out loud that never reach a tool. Output: phase3_omi.md.

Phase 4 — GitHub + code sweep. ~30 min. Tools: gh CLI, local ~/projects/TP3-Neural-Stack git log. Mine commit messages, README voice, PR descriptions. Cross-check Antigravity/Cursor/Jules co-author trailers. Output: phase4_github.md.

Phase 5 — Drive document sweep. ~60 min. Tools: WSL filesystem reads of /mnt/c/Users/Breez/My Drive (...)/AI_Library/. Walk every folder, summarize each top-level doc in one sentence, flag the 10 docs that most reveal who you are vs. what you do. Output: phase5_drive.md.

Phase 6 — Synthesis. ~60 min. Tools: Claude. Merge all phases into Who_Is_Mark_Barnett.md with citation footnotes pointing back to source files. Sections: Identity & background, How he thinks, How he talks, What he's building & why, People in his orbit, Values & non-negotiables, Open loops & anxieties, How to work with him. Hard cap: 3,000 words.

Total est: ~5.5 hours of agent time once armed.

4. Privacy + safety

Final doc lives ONLY at /mnt/c/Users/Breez/My Drive (breezybarnett16@gmail.com)/AI_Library/Who_Is_Mark_Barnett.md and ~/.claude/projects/-home-g16/memory/who_is_mark.md (mirrored). Nowhere else.
Never committed to any git repo, public or private.
Never referenced from any Legacy Soil, Bidet AI, or Saint Francis surface.
Versioned: each run produces Who_Is_Mark_Barnett_v{N}_{date}.md. You review before the canonical file is overwritten.
Hard exclusions: nephews by name, anything Kim would flag, health specifics, legal matters, financial account numbers, anything about the patio life that names another person, any speculation about Shannon/Mom beyond what's already in private memory.
Soft exclusions (ask before including): salary, Saint Francis colleagues by name, conflicts with administrators, any third party who hasn't consented.
Public-gist cleanup recommended: the MrB-Ed/e2511c1740d7359a592709aa4c77a49b gist names you and the architecture. Consider making it secret or deleting after migration is done.
Recommend rotating any tokens that ever touched the public The_Project repo and confirming .env files were never committed there.

5. Two things to fix before Phase 1

1. thebarnetts.info is returning 522 — origin or tunnel is down. Not urgent for research, but worth knowing.

2. The public gist about your secure data management names you and your machines. Recommend deleting or making secret before any deep research goes wider.