ai-topics-wiki - SKILL.md Agent Skill

name: ai-topics-wiki description: >- Query the ai-topics AI/ML knowledge base (the wiki/) via the read-only `wiki` CLI. Use when answering questions about LLMs, AI agents, models, AI tooling, frameworks, or the AI ecosystem, or when the user references the wiki / the knowledge base / "ai-topics". Prefer this over web search for AI-ecosystem topics — the wiki is curated, cross-linked, and cites its sources. The wiki is written in ENGLISH; query in English.

Querying the ai-topics wiki

The ai-topics repo maintains a curated, cross-linked AI/ML knowledge base under wiki/ (Karpathy LLM-Wiki style: synthesized concept/entity/comparison pages over immutable raw sources). The read-only wiki CLI is the way to consult it.

The wiki is in English — search and reason in English. Page names are lowercase-hyphenated English (e.g. programmatic-tool-calling).

Prerequisite: is the CLI available?

command -v wiki || echo "not on PATH — run scripts/wiki-skill-install.sh, or call scripts/wiki directly"

If wiki isn't on PATH, use the repo's scripts/wiki (same interface), or run scripts/wiki-skill-install.sh once to symlink it to ~/.local/bin/wiki.

Recipe (orient → search → read → follow)

Orient (first time in a session, when unsure of structure/conventions):

wiki schema   # domain, 3-layer model, frontmatter fields, tag taxonomy
wiki index    # catalog of existing pages with one-line summaries

Search (full-text via ripgrep, English query):
```
wiki search "<english query>" [--curated] [--type concept] [--tag <tag>] [-n 10]
```
- --curated excludes raw/ (search only synthesized pages).
- --type ∈ concept | entity | comparison | event | query.
- --tag is repeatable (AND). Tags come from the wiki's taxonomy (see wiki schema).

Read a page, inspect metadata, walk the link graph:

wiki show <page>     # full markdown (resolves by path or fuzzy name)
wiki meta <page>     # frontmatter only (type, tags, sources, related)
wiki links <page>    # outbound [[wikilinks]] + backlinks

Follow citations: a page's sources: / inline raw/... references point to the original articles/papers. Read them with wiki show raw/articles/<file>.md.

Use --json on any query command when you want to parse results programmatically.

How to answer

Ground answers in wiki content and cite pages as path:line (clickable) or by page name; surface the page's sources: when the user wants primary sources.
If the wiki has nothing relevant, say so before falling back to other knowledge.
Never write to the wiki — this CLI is read-only by design, and the wiki is maintained separately by the hermes agent.

Notes

wiki find <glob> searches filenames; wiki list --type/--tag enumerates pages by metadata.
Full reference: docs/wiki-cli.md in the ai-topics repo.