me-template

Starter kit for a personal/team knowledge base that an LLM (Claude Code) writes and maintains for you — you curate sources and review diffs, the assistant does the reading/writing. Based on karpathy's LLM wiki pattern.

What this is

A thin scaffold, not a finished system:

• Folder structure + typed YAML frontmatter schema for a wiki of entities (people, systems, decisions, ...)
• Five minimal skills (ingest, query, lint, context-audit, correction-capture) — mechanics only, no opinionated rules baked in
• An optional ## Timeline convention (see CLAUDE.md) for pages that accumulate dated status over time, plus a SessionEnd auto-commit hook (.claude/scripts/vault-autocommit.sh) — both opt-out, not mandates
• An onboard skill that interviews you and turns the scaffold into something that fits your workflow, followed by a usecases skill that turns "it's set up" into 1-2 real things tried, tailored to what onboarding learned about you — say "usecase brainstorm" any time you want that again

It deliberately does not ship a specific person's rules, thresholds, or automation habits (daily briefs, personas, staleness cutoffs, etc.) — those are workflow taste. You build those yourself, if and when you want them, using this as the base layer. One optional, disabled-by-default example is included for illustration: .claude/skills/daily-brief-example/ sketches the shape of a recurring pull-from-an-external-source-then-ingest routine (with _templates/integrations.md explaining the general wiring pattern, using the open-source gws Google Workspace CLI as a worked example). It's not wired into onboarding and safe to delete. A second, equally optional example covers drafting-as-you: me/voice.md + .claude/skills/voice-persona-example/ sketch a voice profile, an outward-facing/private-sparring persona split with an explicit exposure boundary, and a final pass for stripping generic AI-writing tells before a draft goes out — again just a shape to build from, not shipped content.

Use it

git clone https://github.com/toppynl/me-template.git my-vault
cd my-vault
claude

Then run /onboard (or just say "onboard me" / "set up my vault") — it walks you through what to track and how you want to work, and rewrites CLAUDE.md and hot.md to match.

Full pattern explanation, folder layout, and the suggested (editable) entity taxonomy live in CLAUDE.md.

Layout

hot.md         current-context snapshot (you define what "hot" means)
index.md       catalog of every wiki page, one line each
log.md         append-only operation log
me/            who you are, how you want the assistant to work with you
now/           fast-changing current state
wiki/          the knowledge base — people/organisations/systems/... (edit to taste)
.raw/          immutable source material
_templates/    frontmatter skeletons per entity type
.claude/       skills — onboard, usecases, ingest, query, lint, context-audit,
               correction-capture; scripts/ for the optional auto-commit hook
archive/       superseded/obsolete pages