Cc Wiki

cc-wiki — Claude history → Quartz knowledge base

When the user runs /cc-wiki [dir], build a complete Quartz vault from their ~/.claude/ history. Five phases. The first and last are mechanical; the middle three dispatch parallel subagents. End-of-run prints the localhost URL.

Default output dir: ./my-kb. If the user passes [dir], use that. If the directory already exists and is non-empty, ask once before overwriting.

If the user has copied another account's Claude data to a custom location, they can mention it inline (e.g. /cc-wiki --source ~/backup-claude) — pass the same --source <path> to preprocess.py in Phase 1. Otherwise default to ~/.claude.

<SKILL_DIR> resolution: ~/.claude/skills/cc-wiki/ (installed by install.sh). Templates and scripts live under <SKILL_DIR>/templates/ and <SKILL_DIR>/scripts/.

Phase 0 — Scaffold

Bash(mkdir -p <dir>)
Bash(cp -R <SKILL_DIR>/templates/quartz/. <dir>/)
Bash(cp -R <SKILL_DIR>/templates/spine-defaults/. <dir>/)

The Quartz template includes quartz/, quartz.config.ts, quartz.layout.ts, package.json, tsconfig.json, vercel.json, kb.py, and a starter internal/ skeleton. The spine defaults are empty stubs for index.md, methodology.md, arc.md, glossary.md, open-questions.md, disambiguation.md — Phase 4 fills them.

Phase 1 — Pre-process

Bash(python3 <SKILL_DIR>/scripts/preprocess.py --out <dir>/internal [--source <path>])

Reads ~/.claude/history.jsonl + ~/.claude/projects/**/*.jsonl (or <path>/history.jsonl + <path>/projects/**/*.jsonl if --source is set). Writes:

• <dir>/internal/prompts/{project_slug}.md — chronological prompt log per project (chunked by month when > 400 prompts)
• <dir>/internal/sessions/{slug}/{session_id}.md — per-session digest
• <dir>/internal/subagents/{slug}/{session_uuid}/agent-*.md — subagent transcripts
• <dir>/internal/project-overview.{md,json} — authoritative project → corpus mapping (agents use this)
• <dir>/internal/session-inventory.{md,json} — counts and date ranges

If no sessions are found, stop and report: No sessions in ~/.claude/projects/. Nothing to build.

If the user has fewer than ~50 prompts total, warn that the output will be thin but proceed (some signal is better than none).

Phase 2 — Project digests (parallel agents)

Read <dir>/internal/project-overview.json to get the list of projects with prompt counts. Group them:

• Big projects (>200 prompts): one dedicated agent each
• Medium (50-200): one agent each
• Small (<50): bundle 3-5 into a single "cluster" agent

For typical users you'll dispatch 5-12 agents in parallel in a single message. Each agent gets the project-digest prompt template (<SKILL_DIR>/templates/prompts/digest-project.md) with project-specific substitutions: project path, slug, prompt file paths, session/subagent dir paths, and total prompt count.

The template instructs the agent to:

1. Read all assigned prompt files chronologically (not skim)
2. Sample 5-15 session/subagent transcripts for depth
3. Write <dir>/internal/project-digests/{slug}.md with fixed schema (TL;DR → strands → people → companies → tools → concepts → workflow obs → cross-project connections → notable verbatim prompts → open threads)

The template requires verbatim quotes (no paraphrasing) and [[wikilinks]] for every person/company/concept mentioned.

Wait for all agents to finish before Phase 3.

Phase 3 — Cross-cutting synthesis (parallel agents)

Dispatch six agents in parallel, one per axis. Each reads ALL project digests from Phase 2 and writes its assigned output dir. Templates under <SKILL_DIR>/templates/prompts/synthesize-{axis}.md:

Each output dir gets its own index.md plus one page per entry.

Wait for all six to finish before Phase 4.

Phase 4 — Spine (you do this directly)

Write six files with full context of what Phases 2 and 3 produced:

• <dir>/index.md — landing
• <dir>/methodology.md — how this KB was built (cite the pipeline)
• <dir>/arc.md — chronological narrative of the user's six months (or however long)
• <dir>/glossary.md — distilled vocabulary, grouped by category
• <dir>/open-questions.md — unresolved threads, including stated-vs-observed tensions
• <dir>/disambiguation.md — name collisions flagged by Phase 3

Use <SKILL_DIR>/templates/spine-defaults/*.md as starting structure but fill them with the specific evidence from the run. The arc should be a single-read narrative — past tense, dated milestones.

Phase 5 — Build and serve

Bash(cd <dir> && npm install --silent)
Bash(cd <dir> && python3 kb.py build)             # generate .index.json
Bash(cd <dir> && npm run dev)                     # background

npm run dev starts the Quartz server at http://localhost:8080. Run in background and surface the URL.

Final message to the user:

Built {N} pages from {M} prompts. Open http://localhost:8080.

Spine: index, arc, glossary, methodology, open-questions, disambiguation
Content: {p_count} projects, {ppl_count} people, {co_count} companies,
         {c_count} concepts, {w_count} workflows, {t_count} tools, {pb_count} playbooks

⚠️ This vault names real people and unredacted strategy. Don't deploy public without a redact pass.

Hard rules

• One command, no flags. If the user passes [dir], use it. Otherwise default ./my-kb. No other arguments.
• Phases 2 and 3 dispatch in parallel via the Task tool in a single message. Don't fan out sequentially — costs 5-10× more wall time.
• Verbatim quotes only. Every person / decision / pivot in the synthesis must trace back to a quoted prompt from the digest. Forbidden to paraphrase load-bearing claims.
• [[wikilinks]] everywhere. Pre-resolve to lowercase-kebab slugs. Aliases go in frontmatter.
• Don't touch ~/.claude/. Read-only. Pre-processor never writes back.
• No network calls except LLM API. The preprocessor is stdlib; Quartz is local-build.
• If npm install fails: report the error, suggest nvm use 22, and stop. Don't try to repair.
• If a Phase 2 or Phase 3 agent fails: continue with the others, then report which axes have gaps. The KB is still useful with one axis missing.
• Don't write inside <SKILL_DIR>/. Treat it as read-only at runtime; everything goes under <dir>.

Failure modes

• Empty history: No sessions found in ~/.claude/projects/. Stop.
• Single tiny project (<50 prompts): build succeeds but warn output will be thin.
• history.jsonl missing (older Claude Code installs): fall back to session JSONL alone; the corpus is still usable, just missing the project-tagged prompt spine.
• Agent returns malformed digest: re-dispatch that one agent only with Retry with stricter adherence to the digest schema. Don't retry more than once.
• YAML frontmatter error in synthesized page: the page-template forbids unquoted titles with colons / quotes. Phase 3 templates emphasize this.

What this skill does NOT do

• Doesn't pick which projects matter — every project gets analyzed.
• Doesn't push to GitHub or deploy to Vercel.
• Doesn't redact for public use.
• Doesn't run behavioral analysis on the sessions (no session-quality metrics, no model-regression detection).
• Doesn't update across runs — re-running re-builds from scratch. (Future: incremental refresh.)

Source: tejpalv/cc-wiki — distributed by TomeVault.