wiki-query — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited wiki-query (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
Answer a question against an existing Karpathy-style LLM wiki. The pattern this skill follows is described in https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f.
Karpathy's query description, verbatim:
You ask questions against the wiki. The LLM searches for relevant pages, reads them, and synthesizes an answer with citations.
And on filing answers back:
good answers can be filed back into the wiki as new pages. A comparison you asked for, an analysis, a connection you discovered — these are valuable and shouldn't disappear into chat history.
Filing is a judgment call, not an automatic step. Pure fact lookup or restatement of an existing page stays in chat. Cross-page analyses and new connections earn a page.
wiki-init first.wiki_inventory is not callable, tell the user to add the server to their MCP client config and restart.
Read CLAUDE.md. This is the authoritative schema — page categories, frontmatter fields, link conventions, and operation vocabulary. The user may have evolved it since wiki-init. Honor the current state, not the defaults.
wiki_inventory()
No scan_for argument — this is a cheap call that returns every page's slug, frontmatter, body length, and computed in/out links. Combine it with a host Read of index.md to pick candidate slugs. Prefer pages whose title, tags, or category directly match the question. When in doubt, widen the candidate set — reading is cheap.
wiki_read(slug=...)
Pull the full body of each candidate. If a page cites a neighbor ([[other-slug]]) that clearly bears on the question, widen and read it too. Two or three reads are normal; ten means the question is too broad and you should narrow it with the user first.
Write the answer in the primary language declared in CLAUDE.md. Cite every non-trivial claim with [[slug]] pointing at the page the claim came from. Only cite slugs that appeared in the `pages` list returned by step 1 — this is the deterministic guarantee that your citations are not hallucinations. If a claim deserves a citation but no page supports it, say so explicitly ("the wiki does not currently cover X") rather than inventing a slug.
Do not paraphrase so aggressively that the citation loses its anchor. A reader should be able to open the cited page and find the claim within a few lines.
Ask: is this answer worth persisting?
File it back when the answer is:
contain (e.g. "how do X and Y differ on Z?")
conclusion no single page states
Do not file it back when the answer is:
lives on the source page
When filing:
hua-tong-vs-zhen-ding-seasonality, not hua-tong-wins-on-seasonality). Lowercase, dashes, at least two characters.
body to match the schema in CLAUDE.md. The typicalcategory is synthesis. Include frontmatter fields the schema requires, the question as a heading, the cited answer, and a "Sources" section listing every [[slug]] you cited.
wiki_write_page(slug=..., body=..., etag=None). etag=Noneasserts the page does not exist. On conflict, pick a different slug.
index.md using host Read/Write (the MCP tools do not touch the index).
extra_lines=[...]). Use operation="query" unless CLAUDE.md` has defined different vocabulary.
When not filing, skip straight to reporting.
| Step | Tool |
|---|---|
| 1. Scope | wiki_inventory() |
| 1. Scope (index) | host Read on index.md |
| 2. Read candidates | wiki_read(slug=...) |
| 3. Synthesize | (conversation; no tool call) |
| 4. File (page) | wiki_write_page(..., etag=None) |
| 4. File (index) | host Read / Write on index.md |
| 4. File (log) | wiki_log_append(operation="query", ...) |
The llm-wiki-mcp server deliberately owns only page CRUD, inventory, and log append. Index and source fetching stay with the host agent.
When done, report: the answer itself (with citations), the pages read, and whether the answer was filed (and if so, under which slug). Keep the non-answer parts terse — the user can call wiki_inventory for details.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.