add-skill-karpathy-llm-wiki

Add Karpathy LLM Wiki

Sets up a persistent wiki knowledge base on a NanoTars group, based on Karpathy's LLM Wiki pattern. The agent becomes a disciplined wiki maintainer — every source you drop in gets read, integrated into entity/concept pages, cross-linked, and logged. Not RAG retrieval; a living, compounding knowledge base.

Preflight

Before installing, verify NanoTars is set up:

[ -d node_modules ] && echo "DEPS: ok" || echo "DEPS: missing"
docker image inspect nanoclaw-agent:latest &>/dev/null && echo "IMAGE: ok" || echo "IMAGE: not built"
if grep -q "ANTHROPIC_API_KEY\|CLAUDE_CODE_OAUTH_TOKEN" .env 2>/dev/null || [ -f "$HOME/.claude/.credentials.json" ]; then echo "AUTH: ok"; else echo "AUTH: missing"; fi

If any check fails, tell the user to run /nanotars-setup first and stop.

Step 1: Read the pattern

Read ${CLAUDE_PLUGIN_ROOT}/skills/add-skill-karpathy-llm-wiki/llm-wiki.md — Karpathy's full LLM Wiki idea. Understand it before proceeding. Summarise the core idea to the user briefly, then discuss what they want to build.

Step 2: Choose a group

AskUserQuestion: "Which group should host the wiki?"

1. Main group — add to your existing main chat
2. Dedicated group — create a new group just for the wiki (you'll need to register it via /register-group <folder> from your main chat after this)
3. Other — pick an existing group from /list-groups

The wiki ends up at groups/<folder>/wiki/ with raw inputs at groups/<folder>/sources/.

Step 3: Install plugin files

Copy plugin assets into place:

mkdir -p plugins/karpathy
cp -r ${CLAUDE_PLUGIN_ROOT}/files/* plugins/karpathy/

This installs the container-side wiki SKILL at plugins/karpathy/container-skills/wiki/SKILL.md. The plugin loader will mount it into the agent container as /workspace/.claude/skills/wiki/.

Step 4: Design collaboratively

Discuss with the user based on the pattern:

• What's the wiki's domain or topic? (Single-domain or multi-domain registry?)
• What kinds of sources will they add? (URLs, PDFs, images, voice notes, books, transcripts)
• Do they want the full three-layer architecture or a lighter version?
• Any specific conventions they care about? (The pattern intentionally leaves this open.)

Step 5: Scaffold the wiki

Based on the discussion, create the wiki structure inside the chosen group folder:

mkdir -p groups/<folder>/wiki/{entities,concepts,reading-notes}
mkdir -p groups/<folder>/sources

Initial files:

cat > groups/<folder>/wiki/index.md <<'EOF'
# Wiki Index

Catalog of every wiki page. Read first on every query. Update on every ingest.

## Entities

(none yet)

## Concepts

(none yet)

## Reading notes

(none yet)
EOF

cat > groups/<folder>/wiki/log.md <<'EOF'
# Wiki Log

Append-only chronological record of ingests, queries, lints.

## [<TODAY>] init | wiki created
EOF

For multi-domain wikis, also create wiki/_registry.md cataloguing each domain subdirectory and add a per-domain SCHEMA.md template.

Step 6: Wire the group's CLAUDE.md

Edit groups/<folder>/CLAUDE.md (or its .local.md companion) to point at the wiki and the container skill. Add a section like:

## Wiki

You maintain a persistent personal knowledge wiki for the user.

- `wiki/` — your maintained markdown pages (entities, concepts, reading-notes, syntheses)
- `sources/` — raw curated inputs the user gives you (read-only)
- `wiki/index.md` — catalog; read FIRST on every query
- `wiki/log.md` — append-only chronological record
- `/workspace/.claude/skills/wiki/SKILL.md` — full schema and ingest discipline. Read it when in doubt.

**Three operations:** `ingest` (a new source arrives), `query` (a question to answer from accumulated knowledge), `lint` (periodic health check).

**Ingest discipline (critical):** when the user provides multiple sources or a folder, process them strictly **one at a time**. For each: read in full → discuss takeaways → update wiki pages → update index + log → THEN move to the next. Never batch-read and synthesise — that produces shallow generic pages.

Step 7: Optional weekly lint schedule

AskUserQuestion: "Schedule a periodic wiki health check?"

1. Weekly (Sun 10:00 local) — recommended
2. Monthly (1st of month, 10:00)
3. Skip — lint manually

If yes, schedule a NanoTars task. Use the /register-group flow to confirm the group is wired, then ask the user to run /list-groups so we have the chat_jid + folder. Insert via:

node -e "
const Database = require('better-sqlite3');
const { CronExpressionParser } = require('cron-parser');
const db = new Database('store/messages.db');
const cronExpr = '0 10 * * 0';   // weekly Sun 10:00
const interval = CronExpressionParser.parse(cronExpr, { tz: process.env.TZ || 'UTC' });
const nextRun = interval.next().toISOString();
db.prepare('INSERT INTO scheduled_tasks (id, group_folder, chat_jid, prompt, schedule_type, schedule_value, context_mode, next_run, status, created_at) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)').run(
  'wiki-lint',
  '<group_folder>',
  '<chat_jid>',
  'Run a wiki health check per /workspace/.claude/skills/wiki/SKILL.md (the LINT operation). Walk wiki/ and report contradictions, stale claims, orphan pages, missing entity/concept pages, missing cross-references, and data gaps. Report findings as a compact list. Do NOT auto-fix — the user decides. Append a YYYY-MM-DD lint entry to wiki/log.md.',
  'cron',
  cronExpr,
  'group',
  nextRun,
  'active',
  new Date().toISOString()
);
db.close();
console.log('wiki-lint task created');
"

For monthly: 0 10 1 * * instead.

Step 8: Source-handling capabilities

Based on the source types the user plans to ingest (Step 4), confirm needed capabilities are installed:

• URLs → built-in WebFetch returns summaries; for full-text ingestion the agent should curl directly into sources/ (the container skill specifies this).
• PDFs → may need a PDF reader skill if not already present.
• Voice notes → install transcription via /add-skill-transcription if not present.
• Images / vision → handled natively by Claude.

If a needed capability is missing, point the user at the relevant /add-skill-* flow.

Step 9: Restart and test

nanotars restart

Test: send a URL or short article to the wiki group. The agent should fetch it into sources/, discuss takeaways, then write reading notes + entity/concept pages and update index + log.

Verify

• ls groups/<folder>/wiki/ — should show index.md, log.md, plus subdirs
• After first ingest: cat groups/<folder>/wiki/log.md — should show the timestamped ingest entry
• tail -10 logs/nanotars.log | grep -i wiki — agent should reference the wiki skill

Uninstall

1. nanotars stop
2. rm -rf plugins/karpathy/
3. The wiki content under groups/<folder>/wiki/ is preserved — delete manually if wanted: rm -rf groups/<folder>/wiki groups/<folder>/sources
4. Remove any wiki-lint scheduled task via /delete-group <folder> and re-register without it, OR file a backlog request for a /delete-task admin command (not yet shipped).
5. nanotars restart