LLM Wiki Manager
The LLM writes, cross-references and keeps the books while you curate sources — yielding a persistent, compounding personal wiki instead of a rotting notes folder.
SKILL.md defines a clear least-privilege layering: the LLM never writes raw/, migrate defaults to dry-run, multi-wiki routing requires an explicit CLAUDE.md declaration, contradictions are flagged rather than overwritten, and every claim cites back to raw/ — transparent data flow and git-based rollback. Deducted: bulk writes to wiki/ and multi-wiki promote lack a uniform explicit confirmation gate (only Update mode has per-page diff confirmation), and script sources are outside this path's evidence, so isolation/recovery claims cannot be verified from the skill file alone.
Instruction self-consistency is strong: nine core invariants, a mode router, Step 0 schema version checks, idempotent script conventions, and failure feedback (lint exit code 1, migrate dry-run) are clearly specified. Deducted: static review cannot execute anything; scripts/ sources and tests are not in this path's evidence, so idempotence and error handling are unverified and the executed ceiling applies.
Audience and scenarios are clear (personal research wiki / second brain), triggers are concretized in a mode-routing table, Update vs Ingest boundaries get dedicated treatment, and the skill depends only on filesystem + git (no unreachable overseas services; mainland-China network impact minimal). Deducted: non-fit boundaries (wiki size limits, non-markdown sources, when NOT to trigger) are under-specified, and there is no Chinese-language documentation or trigger support.
Good progressive disclosure (SKILL.md → references → templates), MIT license, semantic versioning, a thorough CHANGELOG, bilingual READMEs, and release CI are all visible. Deducted: maintenance responsibility and update path rest only on an individual repo, and referenced files (references/, scripts/, templates/) were not directly auditable within this path, leaving hidden-assumption risk.
The value proposition is clear with a comparative argument (compiled wiki vs re-reading via RAG), and outputs (markdown pages, lint reports, logs) are in a directly usable format. Deducted: static review cannot verify key-path outputs; the nutrition example is illustrative rather than validated; cost/benefit (LLM context cost of reading sources) is unquantified — capped at the static execution limit.
Auditable primary material exists: versioned CHANGELOG, a CI release workflow, and external corroboration via Karpathy's gist. Deducted: no committed test suite, no independent reproduction of script behavior, and functional claims rest mainly on author statements — only the low band is awardable statically.
- Static review only: no scripts were executed; all scores derive from source reading at low confidence.
- scripts/ sources were not directly auditable within this assessment path; idempotence, error handling, and lint behavior are author claims — review them yourself and dry-run in a test directory first.
- Ingest and multi-wiki promote perform bulk writes across many files; use inside a git repo for rollback and watch for pre-write confirmation prompts.
- Documentation and triggers are English-only; Chinese users must phrase triggers in English or adapt workflows manually.
- Publisher is unverified by the FollowSkills registry; identity is unknown (not itself a deduction), so monitor repo activity for maintenance sustainability.
What does this skill do, and when should you use it?
A Claude Code skill implementing Karpathy's LLM Wiki pattern. Rather than re-deriving answers from raw documents on every query like RAG, it compiles your sources into a persistent, interlinked markdown knowledge base. Each ingest updates entity pages, concept pages, the index and an operation log, and eagerly flags contradictions. It ships 5 idempotent Python scripts, 8 operating modes, 8 templates and 11 reference docs. The core discipline: the LLM does all the bookkeeping.
Runs init_wiki.py to scaffold raw/, wiki/ and CLAUDE.md in your research folder; on ingest, reads PDFs/articles/transcripts, writes summary pages and updates cross-references on entity and concept pages; maintains a category-organized index.md via update_index.py and a greppable operation log via append_log.py; fully rewrites the ~500-word hot.md hot cache after every ingest; update mode performs a semantic sweep across pages carrying a superseded claim, with diff-before-write per page; lint_wiki.py checks for orphan pages, broken links (including [[wiki-links]]), index drift and hot.md bloat, auto-writing dated reports to wiki/reports/; migrate_wiki.py handles v1→v2 schema upgrades with a dry-run; multi-wiki mode routes writes between a project wiki and a global wiki (e.g. an Obsidian vault) based on an External Wiki declaration in the project's CLAUDE.md.
- A researcher or long-horizon investigator building a 'second brain' who wants every paper and article to become an interlinked knowledge page rather than a loose file
- Someone with 50+ documents who currently has Claude re-read the corpus on every question and wants a pre-compiled wiki to answer from
- An Obsidian user with a global vault across projects who wants to promote portable concepts to the global wiki while keeping project knowledge local
- Anyone tracking a fast-moving field who needs to correct one claim paraphrased across many pages when new research supersedes it
- A Claude Code developer who wants the same pattern for project-specific decisions, architecture notes and research
What are this skill's strengths and limitations?
- All five scripts are idempotent and depend only on the Python standard library — zero pip dependencies
- Clean three-layer model: raw/ is read-only so the audit trail holds, wiki/ is fully LLM-owned, CLAUDE.md carries an evolving schema
- Contradictions are flagged side by side with sources, never silently overwritten — the user decides
- Lint reports are dated and auto-archived; git log wiki/reports/ shows wiki quality over time
- Plain markdown + git: readable in Obsidian, VS Code or vim, no vendor lock-in
- Very complete reference material (11 workflow docs, templates, a mature-wiki example)
- Mode selection is fully inferred from natural language with at most one clarifying question — misrouting is possible
- Update mode's multi-page semantic sweep relies on model judgment; paraphrases grep can't find may still be missed, and no test suite is shown
- Non-Claude-Code platforms (Copilot, Cursor, Codex, Gemini CLI) are listed with skill paths in the README but unverified
- Value depends on ongoing ingestion and discipline — skipping log/index updates lets the wiki rot, and the skill cannot fully prevent that
- Schema versioning and some workflows add ceremony that may feel heavy for a small note collection
How do you install this skill?
Any of: 1) git clone https://github.com/sametbrr/llm-wiki-manager ~/.claude/skills/llm-wiki-manager; 2) gh skill install sametbrr/llm-wiki-manager (requires gh CLI v2.90+); 3) download the .skill file from releases and unzip into ~/.claude/skills/llm-wiki-manager. Requires Claude Code or any agentskills.io-compatible agent and Python 3.9+ (stdlib only, no pip install). Start a new Claude Code session after installing.
How do you use this skill?
Launch claude in your research folder and trigger modes with natural language — no slash commands. Example flow: 'Set up an LLM wiki here. Topic: history of nutrition science.' (bootstrap) → drop a PDF into raw/ → 'Ingest Pollan's In Defense of Food' (ingest) → 'What does the wiki say about nutritionism?' (query) → 'Lint the wiki' (health check, report auto-saved to wiki/reports/). Other triggers: 'Smith 2024 supersedes Keys 1980' (update), 'Add this to my global wiki' (multi-wiki), 'We should always tag book chapters' (schema-evolve).
How does this skill compare with similar options?
The explicit contrast is with RAG: RAG re-discovers answers from raw documents on every query, while this pattern compiles sources once into an interlinked persistent wiki that queries read directly, with contradictions already flagged. The pattern originates from Karpathy's LLM Wiki gist; this repo is a full Claude Code skill implementation of it.