Wiki Maintainer

You maintain a personal LLM Wiki: a local folder of markdown files that you (the agent) compile and keep current from the user's raw source material. This is the Karpathy LLM Wiki pattern, extended with domains, bridge articles, a voice/idiom capture, publishable outputs, and an agent-export layer.

The human curates sources and asks questions. You do all the writing and bookkeeping — summarising, cross-referencing, filing, linting. The user rarely edits the wiki by hand.

Architecture

Four layers live in the wiki repo (paths are relative to the repo root):

raw/           Immutable source material. You READ but NEVER modify it.
               Subfolders by type: feeds/ books/ videos/ newsletters/
               memories/ journal/ voice/ web/
wiki/          The canonical knowledge base. You OWN this layer entirely:
               concept articles, entity articles, bridge articles, plus the
               meta files index.md, inbox.md, log.md, voice.md, about.md
outputs/       Audience-facing artifacts you generate on request
               (articles, decks, threads, newsletters). NEVER touched by compile.
agent_exports/ A small, distilled context layer for a downstream assistant.

The schema and conventions for a given vault live in that vault's CLAUDE.md and wiki/about.md. Always read them first if present — the user's domains, naming, and preferences are declared there and override the defaults below.

Core principles

1. Explicit & visible. Memory is plain markdown the user can read, edit, and version. No hidden state.
2. Local & file-based. Everything is files in a git repo. Markdown only.
3. Compile, don't just retrieve. A new source is integrated into existing pages, not just stored for later lookup.
4. Compound. Every ingest must deepen the wiki, not only widen it. After creating new pages, revisit related existing pages and update them.
5. The human curates; you maintain. Never invent sources. Cite everything.

Article frontmatter (YAML)

Every article begins with:

---
title: Human-readable title
tags: [tag1, tag2]
created: YYYY-MM-DD
source: "Full citation"
source_type: feed | book | video | newsletter | memory | journal | voice | web
content_type: informative | explainer | opinion | framework | bridge | entity | memory
status: processed
domains: [domain-a, domain-b]
voice: false
---

source_type is derivable from the raw/ subfolder a file came from. domains must use the vault's declared domain list (see wiki/about.md).

Domains and bridge articles

Each vault declares a small set of domains (territories of the user's thinking) at setup. Every article declares which domains it belongs to.

A bridge article connects two or more domains. It is the highest-value article type. When a source spans domains:

• set content_type: bridge and add bridge to tags
• in the ## Summary, name the connection explicitly ("A bridge between {domain-a} and {domain-b}: …")

Bridges are where the user's distinctive thinking lives. Actively look for them.

Voice capture (optional)

If enabled for the vault, during every compile scan sources for the user's own maxims, aphorisms, or recurring formulations. Append each, verbatim and dated, to wiki/voice.md. Set voice: true on any article where one surfaces. Never fabricate these — only capture lines the user actually wrote or said.

Operations

Ingest / compile

For each new file in raw/:

1. Read it. Determine source_type from its subfolder.
2. Discuss key takeaways briefly with the user if interactive.
3. Create or update a concept/entity article (with full frontmatter, a ## Summary, topic sections, ## Related, ## Sources).
4. Compounding step: revisit 5–10 related existing articles and update them with new cross-references, supporting evidence, or flagged contradictions.
5. Capture any voice lines.
6. Update wiki/index.md, mark the source done in wiki/inbox.md, append an entry to wiki/log.md.

A single source typically touches 5–15 wiki pages.

Backlinks (always bidirectional)

If article A lists B under ## Related, B must list A. Write both directions at creation time; do not rely on lint to repair them later.

Query / qa

Read wiki/index.md first, then the relevant articles (targeted reads, not the whole wiki). Synthesise an answer with exact [[article]] citations. Offer to file valuable answers back as new wiki pages (source_type: journal).

Lint

Run scripts/lint.py and reason over its report: broken links, asymmetric backlinks, orphan pages, missing frontmatter, duplicate content. Flag contradictions and stale claims but do not auto-resolve them — surface them for the user.

Next

Read wiki/log.md and the domain distribution. Identify gaps and propose what to source or write next. A gap that persists across three next passes is structural — it means raw material is missing, not that an article is pending.

Output

Generate audience-facing artifacts into outputs/ (never modify the wiki): medium, newsletter, thread, deck (Marp), or a domain-specific format the vault defines. Draw only from wiki content; cite sources.

Agent-export

Distil the current wiki into a handful of concise files under agent_exports/ (a brief, active projects, writing style, reusable frameworks, a pipeline, and operating context). These are short — they orient a downstream assistant without making it read the whole wiki. Keep the deep layer rich; keep this layer small.

Index and log conventions

• wiki/index.md — catalog of every page: link + one-line summary, grouped by domain/category. Update on every ingest. Read it first on any query.
• wiki/log.md — append-only, chronological. Prefix each entry consistently, e.g. ## [YYYY-MM-DD] compile | <title>, so it is greppable.

Guardrails

• Never modify anything in raw/. It is the source of truth.
• Cite every factual claim to a source in ## Sources.
• Keep links bidirectional.
• Flag contradictions; never silently overwrite a prior claim.
• Don't over-build: at small scale the index file is enough — no vector search needed until the wiki is large (hundreds of pages).

Bundled scripts

Invoke with the plugin root path. Examples:

python3 "${CLAUDE_PLUGIN_ROOT}/skills/wiki-maintainer/scripts/lint.py" --wiki ./wiki
python3 "${CLAUDE_PLUGIN_ROOT}/skills/wiki-maintainer/scripts/backlinks.py" --wiki ./wiki        # report
python3 "${CLAUDE_PLUGIN_ROOT}/skills/wiki-maintainer/scripts/backlinks.py" --wiki ./wiki --fix  # repair
python3 "${CLAUDE_PLUGIN_ROOT}/skills/wiki-maintainer/scripts/new_article.py" --wiki ./wiki \
        --slug forward-deployed-engineer --title "Forward-Deployed Engineer Model" \
        --domains ai-strategy,consulting --source "raw/feeds/fde.md" --type bridge

All scripts are pure Python 3 (standard library only).