garrytan/gbrain · skills/

The gbrain Skill Atlas

53 markdown skill files, explained as one story: what happens to a message from the moment it arrives to the moment the brain dreams about it at 3 AM.

A skill is just a markdown file with instructions. The agent reads the file, then does what it says. No code to run — the file is the program. Ten steps below, one diagram each. Every skill name below links to its file on GitHub, so you can read the real thing in place.

writes  changes the brain reads  only looks things up
1

How a skill gets picked

Every message first hits RESOLVER.md — a phone book of trigger phrases. Match a phrase, read that skill file, follow it. If two match, read both: skills are built to chain.

your message "save this thought" RESOLVER.md trigger → skill path SKILL.md read it, then act gbrain CLI / MCP matches points to calls two matches? read both — skills chain Rule #1 in the resolver: “Read the skill file before acting.”
The dispatcher. Nothing clever happens here — the intelligence lives in the skill files it points to.
2

The two skills that never sleep

Two skills run on every message, no trigger phrase needed. Together they make the core loop: notice → look up → answer → save → link → sync.

  • signal-detectorwritesRuns in the background on every inbound message. Quietly grabs names, ideas, todos, links — before you even ask.
  • brain-opswritesThe house rules for any brain read/write: look in the brain FIRST (before the web), cite sources, always back-link.
signal search respond write auto-link sync every message brain first with context page + timeline zero LLM calls cron keeps fresh every write feeds the next search — the loop is why the brain compounds
The loop from the README. Amber boxes mutate the brain; auto-link builds the typed graph on every write, no LLM needed.
3

Getting stuff in

One front door (ingest) that looks at what you shared and hands it to the right specialist. Everything lands as a markdown page.

you share or a webhook fires ingest idea-ingest media-ingest meeting-ingestion voice-note-ingest capture brain page inbox/2026-08-07-… links · tweets · articles video · PDF · book · repo transcripts + attendees exact words, never paraphrased raw thought, one command detects type, delegates
The intake fan. Each specialist knows its format's quirks; all of them end at the same place — a markdown page with frontmatter and links.
  • capturewritesOne command to save any thought. Replaces "which of three save paths do I use?" with gbrain capture "…".
  • ingestwritesThe router above. Looks at input type, delegates.
  • idea-ingestwritesLinks, articles, tweets, ideas. Fetches content, saves it, cross-links it.
  • media-ingestwritesYouTube, podcasts, PDFs, whole books, screenshots, GitHub repos.
  • meeting-ingestionwritesMeeting transcripts → pages, plus enrichment for every attendee.
  • voice-note-ingestwritesVoice notes, with your exact phrasing preserved verbatim.
  • webhook-transformswritesTurns outside events (SMS, Zapier, calendar) into brain signals.
  • archive-crawlerwritesDigs through old archives (Dropbox, Gmail takeout, hard drives) for your own writing and relationships. Refuses to bulk-dump junk in.
  • data-researchwritesStructured extraction: pull tables of facts (investor updates, donations) out of sources into pages.
4

Filing it in the right place

A brain where everything lands in inbox/ forever is just a pile. Three gate skills keep the shape; one sweeps up after big work sessions.

new page from any ingest brain-taxonomist which folder does this live in? frontmatter-guard is the YAML header valid? write it people/ companies/ … after a big work session, eiirp ("Everything In Its Right Place") sweeps loose files back through the gate
The filing gate. Consulted before any new page is created — misfiled pages are what make old wikis rot.
  • repo-architecturereadsThe rulebook: where each kind of file goes, by its primary subject.
  • brain-taxonomistreadsThe gate that applies the rulebook before every new page.
  • frontmatter-guardwritesValidates and auto-repairs the YAML header on pages.
  • eiirpwritesPost-work organizer. "File all of this" — archives a whole research thread properly.
  • schema-authorwritesEvolves the brain's shape itself: adds new page types (e.g. lab-result) when your data outgrows the defaults.
  • schema-unifywritesMigrates a messy 94-type brain down to the 15 canonical types.
5

Getting answers out

Asking is three searches at once, fused, then written up — with citations and an honest "here's what I don't know."

question "who works at Acme?" vector meaning-alike chunks keyword exact words (BM25) graph typed edges: works_at… fuse rank + rerank the answer prose + [citations] + "the brain hasn't heard about Alice in 6 weeks" synthesize
The query path (gbrain search stops at "fuse"; gbrain think goes all the way). The gap note in green is the differentiator.
  • queryreadsThe skill that runs the picture above: 3-layer search, synthesis, citations.
  • idea-lineagereadsTraces one idea through time: first mention → how it evolved.
  • strategic-readingwritesReads a book/article through the lens of a problem you're facing; outputs an applied playbook.
  • book-mirrorwritesChapter-by-chapter book analysis mapped to your life, using brain context.
  • concept-synthesiswritesSquashes thousands of raw concept stubs into a tiered intellectual map (Canon → Riff).
  • publishreadsShares a brain page as a password-protected HTML link.
  • brain-pdfreadsTurns a brain page into a clean PDF. Page stays the source of truth.
6

Making thin pages rich

A person page with one line is a stub. Enrichment asks the web only for what the brain doesn't already know, writes it back with citations, then audits the citations.

thin page people/alice: 1 line enrich tiered protocol perplexity-research web: only what's NEW vs brain write back + cite citation-fixer audits later the dashed loop: enrichment re-runs on a schedule, so pages keep pace with reality
The enrichment chain. Brain context goes out with the web query, so you never pay to re-learn what you already know.
  • enrichwritesFills out person and company pages, tier by tier.
  • perplexity-researchwritesBrain-aware web research: returns what's new vs what's known.
  • academic-verifywritesChases a research claim to its source: paper → method → data → replication.
  • article-enrichmentwritesTurns raw article dumps into structured pages: summary, quotes, why-it-matters.
  • citation-fixerwritesAudits pages for facts with no [Source: …], fixes formats.
7

The night shift: dream + maintain

While you sleep, gbrain dream runs eight phases. The two in the middle are the memory-consolidation part: yesterday's conversations become permanent, linked pages.

lintbacklinkssync synthesizeextractpatterns embedorphans yesterday's chats medical/therapy excluded first cheap filter (Haiku) "worth keeping?" — most aren't 1 subagent per keeper quotes you verbatim reflections + ideas allow-listed folders only 30 days of reflections short-term memory "same theme ≥ 3 times?" one Sonnet pass pattern page long-term memory guardrails: cooldown = max ~2 runs/day · re-running the same transcript is a no-op · injection-proof write allow-list this is the "pruning": junk is filtered at the door, and repetition is squashed into patterns — pages are almost never deleted
The 8-phase dream cycle, with the two consolidation phases expanded. Amber = the phases that create new memory.

And when something's broken: the doctor loop

score the brain gbrain doctor → 71/100 make a plan ordered steps, $ + time est. run one step as a background job re-score better? loop until: target score hit ✓ · or --max-usd cap hit ✕ · or target impossible (it says why and stops) gbrain doctor --remediate --yes --target-score 90 --max-usd 5
Self-repair with a hard dollar cap, so an unattended fix loop can't burn your API credits.
  • maintainwritesThe 13-dimension health audit: stale pages, orphans, dead links, missing citations, tag mess, stale embeddings…
  • smoke-testwrites"Did the restart break anything?" Post-restart checks + auto-fix.
  • skillpack-checkreadsHealth report for the skill files themselves.
  • cron-schedulerwritesManages the overnight schedule: staggering, quiet hours.
  • minion-orchestratorwritesThe job queue driver: background subagents that survive crashes.
  • cross-modal-reviewreadsSecond opinion from a different AI model before trusting big work.
  • gbrain-upgradewritesKeeps gbrain itself current when a version nag appears.
8

The daily rhythm

daily-task-prep briefing daily-task-manager dream + maintain calendar + meeting context "what's happening today" add · complete · defer consolidate while you sleep 7 am8 amall day3 am
Ops skills mapped to a day. cron-scheduler places all of these; quiet hours keep the agent from pinging you at 3 AM.
  • daily-task-prepreadsMorning: looks ahead at your calendar, loads context for each meeting, lists open threads.
  • briefingreadsThe daily digest: meetings, active deals, what changed.
  • daily-task-managerwritesTask lifecycle: add, complete, defer, review.
  • reportswritesSaves timestamped reports so cron jobs and humans can find them later.
  • ask-userreadsThe pattern for "stop and give the human 2–3 clear options" before big moves.
  • gbrain-advisorreadsWeekly coach: "here's what your brain setup is missing."
9

Skills that make skills

The meta layer. A trick you did once becomes a file; the file gets tested, optimized against a benchmark, and shipped upstream so every gbrain user's agent learns it.

a good trick you did it once, by hand skillify → skill-creator write SKILL.md properly testing does it validate? skill-optimizer keep only edits that score higher skillpack-harvest lift it upstream to gbrain benchmark loop …and back down: every gbrain install gets the skill on next upgrade
The skill lifecycle. This loop is why the README says "as my personal agent gets smarter, so does yours."
  • skillifywrites"Is this a skill?" Turns a raw capability into a proper, tested one.
  • skill-creatorwritesGenerates a conformant SKILL.md scaffold.
  • skill-optimizerwritesTreats the skill text as a trainable parameter; benchmark-gated edits.
  • testingreadsValidates skills + watches the test suite for regressions.
  • skillpack-harvestwritesPromotes a proven local skill upstream into the shared pack.
10

Day zero: setup & migration

  • setupwritesFirst boot: makes the database (2-second local PGLite or Supabase), wires the agent, first import.
  • installwritesThe agent-readable install protocol (what INSTALL_FOR_AGENTS.md drives).
  • cold-startwrites"Now what?" Sequences the highest-value first imports for an empty brain.
  • migratewritesMoves you in from Obsidian, Notion, Logseq, Roam, CSV, JSON…
  • soul-auditwritesA 6-phase interview that writes the agent's identity file (SOUL.md) — who it is, how it talks to you.

That's all 53. The README says "43 curated skills" — the pack has grown since that sentence was written.

Full roster, one line each

SkillIn one sentence
Always on
signal-detectorCatches names, ideas, todos in every message, silently.writes
brain-opsHouse rules for every brain read/write: brain first, cite, back-link.writes
Intake
captureSave any thought with one command.writes
ingestFront door; routes to the right specialist.writes
idea-ingestLinks, tweets, articles, ideas.writes
media-ingestVideo, audio, PDF, books, screenshots, repos.writes
meeting-ingestionTranscripts → pages + attendee enrichment.writes
voice-note-ingestVoice notes, exact words kept.writes
webhook-transformsOutside events (SMS, Zapier) → brain signals.writes
archive-crawlerMines old archives for your own high-value writing.writes
data-researchExtracts structured data (tables of facts) from sources.writes
Filing & shape
repo-architectureThe rulebook for where files go.reads
brain-taxonomistThe gate that applies the rulebook before every write.reads
frontmatter-guardFixes broken YAML headers.writes
eiirpPost-work sweep: everything in its right place.writes
schema-authorAdds new page types when your data needs them.writes
schema-unifyCollapses type sprawl to the 15 canonical types.writes
Answers out
query3-layer search + synthesized, cited answer.reads
idea-lineageTraces one idea's evolution over time.reads
strategic-readingReads a source through the lens of your current problem.writes
book-mirrorBook analysis mapped to your actual life.writes
concept-synthesisThousands of concept stubs → tiered idea map.writes
publishShare a page as protected HTML.reads
brain-pdfPage → clean PDF.reads
Enrichment
enrichFills out person/company pages, tiered.writes
perplexity-researchWeb research that only returns what's new vs the brain.writes
academic-verifyTraces claims to paper → data → replication.writes
article-enrichmentWall-of-text dumps → structured, quotable pages.writes
citation-fixerEvery fact gets a source, formatted right.writes
Health & night shift
maintain13-dimension audit + the dream cycle runbook.writes
smoke-testPost-restart checks with auto-fix.writes
skillpack-checkHealth report for the skills themselves.reads
cron-schedulerStaggered schedules, quiet hours.writes
minion-orchestratorCrash-proof background jobs and subagents.writes
cross-modal-reviewSecond opinion from a different model.reads
gbrain-upgradeKeeps gbrain itself up to date.writes
Daily ops
daily-task-prepMorning calendar lookahead + meeting context.reads
briefingThe daily digest.reads
daily-task-managerAdd, complete, defer, review tasks.writes
reportsTimestamped, findable reports.writes
ask-userStop and offer the human clear choices.reads
gbrain-advisorWeekly "get more out of your brain" coaching.reads
Meta (skills about skills)
skillifyRaw capability → proper skill.writes
skill-creatorConformant SKILL.md scaffolds.writes
skill-optimizerBenchmark-gated skill text optimization.writes
testingSkill validation + regression watch.reads
skillpack-harvestPromote a proven skill upstream.writes
functional-area-resolverShrinks a bloated routing file.writes
Day zero
setupFirst boot: DB, agent wiring, first import.writes
installThe agent-driven install protocol.writes
cold-startWhat to import first into an empty brain.writes
migrateIn from Obsidian / Notion / Logseq / Roam / CSV.writes
soul-auditInterview that writes the agent's SOUL.md identity.writes

Source: skills/ in garrytan/gbrain, read 2026-08-07. Write/read flags from each skill's mutating: frontmatter (judgment call where unmarked).

← Accession