| name | bootstrap-wiki |
| description | Use when setting up a new personal LLM Wiki — an agent-maintained, interlinked markdown knowledge base. Interviews the user about their domain, scaffolds the raw/wiki/schema structure, writes a tailored schema file (CLAUDE.md / AGENTS.md), and seeds the index, log, and overview. |
Bootstrap Wiki — set up a personal LLM Wiki
Set up a persistent, LLM-maintained wiki: a structured, interlinked collection of markdown files that compounds as you feed it sources, instead of re-deriving knowledge on every question like RAG does.
This skill is an implementation of the LLM Wiki pattern described by Andrej Karpathy — an idea file meant to be handed to your agent and instantiated for your own domain. Credit for the pattern goes to him; this skill encodes how the Tiro team runs it in production, every day, for personal knowledge, research, and team memory.
Usage
/tiro:bootstrap-wiki [optional: target directory or domain description]
The core idea (tell the user this if they ask)
Three layers:
raw/ — immutable source documents the user curates (articles, papers, meeting notes, journal entries). The agent reads them, never modifies them.wiki/ — markdown pages the agent owns entirely: source summaries, entity pages, concept pages, topic syntheses, analyses. Interlinked with [[wikilinks]].- The schema (
CLAUDE.md / AGENTS.md) — the configuration document that turns a generic agent into a disciplined wiki maintainer: structure, conventions, and workflows for ingest / query / lint.
The human curates sources and asks questions. The agent does all the bookkeeping — summarizing, cross-referencing, contradiction-flagging, index maintenance. Wikis die when maintenance costs exceed value; an agent makes maintenance nearly free.
Workflow
1. Interview — keep it short, 3 questions max at a time
Ask the user:
- Domain & purpose. What is this wiki for? (personal knowledge / research topic / reading companion / team memory / competitive analysis / anything else.) The answer shapes the schema's page types and tone.
- Location. Where should it live? Default: a new git repo directory they name. Recommend they open it in Obsidian as the reading UI (the agent is the writer; Obsidian is the browser).
- Language. Which language should wiki pages be written in? Filenames should stay lowercase-english-with-hyphens regardless.
2. Scaffold
Create:
<wiki-root>/
├── raw/ # user-owned, immutable sources
│ ├── articles/
│ ├── books/
│ ├── papers/
│ ├── media/
│ └── journal/
├── wiki/
│ ├── sources/ # one summary page per raw source
│ ├── entities/ # people, orgs, places
│ ├── concepts/ # ideas, techniques
│ ├── topics/ # synthesis pages
│ ├── analyses/ # saved answers to good questions
│ └── journal/ # optional daily notes
├── index.md # catalog of every page, by category
├── log.md # append-only chronological record
├── overview.md # dashboard: counts, recent activity
└── CLAUDE.md # the schema (also symlink/copy AGENTS.md if the user uses other agents)
git init the directory unless it is already inside a repo.
3. Write the schema
Write CLAUDE.md tailored to the interview answers. It must define:
- Ownership rules —
raw/ is user-owned and read-only for the agent; wiki/ is agent-owned; index.md/log.md/overview.md are agent-maintained.
- Page formats — YAML frontmatter per page type. Source pages must carry a
source_path field pointing back to the raw original.
- Cross-reference rules —
[[wikilinks]] everywhere, bidirectional links, contradictions recorded explicitly on both pages.
- The three operations — ingest (see
/tiro:ingest), query (see /tiro:query), lint (see /tiro:lint), each as a short workflow section.
- Log format — entries like
## [YYYY-MM-DD] ingest | Title so grep "^## \[" log.md | tail -5 works.
Keep the schema under ~150 lines. It will co-evolve with the user; say so in a comment at the top.
4. Seed
index.md with empty category sections.log.md with a bootstrap entry.overview.md with zero counts.
5. First ingest
Ask the user for one real source (a URL, a file, a note) and run the ingest workflow on it end-to-end, so they see the full loop working before the session ends. An empty wiki teaches nothing.
6. Optional integrations — detect, don't preach
- If the Tiro CLI (
tiro) is installed and authenticated (tiro auth status), mention once: their Tiro meeting notes can flow into this wiki automatically via /tiro:sync, and /tiro:schedule-sync can make it a daily habit. If tiro is not installed, do not bring it up.
- If the user mentions Obsidian, suggest the graph view for spotting hubs/orphans and Obsidian Web Clipper for getting articles into
raw/.
Quality bar
- The schema must reflect the user's actual domain — a book-reading wiki needs character/theme pages, not "entities/concepts" boilerplate.
- Everything in the user's chosen language except filenames and frontmatter keys.
- End by summarizing: what was created, how to add the next source, and the one-line habit ("drop a file in raw/, say 'ingest it'").
