arthai-marketplace

/wiki-knowledge-base

Build and maintain LLM-powered topic wikis.

Synopsis

/wiki-knowledge-base <init|fetch|ingest|query|lint|status> <topic> [-- <text>]

When to use it

Researching a technology or vendor landscape before adopting it
Competitive analysis or domain deep-dives where knowledge accumulates over weeks
Building durable domain context that /planning, /fix, /implement, and /review-pr automatically pick up
Not for: project code knowledge — that lives in the /calibrate knowledge base (calibrate)

Quickstart

/wiki-knowledge-base init stripe-payments

What you’ll see: the skill asks you to describe the wiki’s focus in 1-2 sentences, then scaffolds .claude/wikis/stripe-payments/ with raw/ (for your source documents), raw/assets/ (for images/PDFs), wiki/ (index, log, pages), and a schema.md. Add source documents to raw/ (or use fetch), then run ingest.

Examples

/wiki-knowledge-base init video-research                                    # scaffold a new wiki
/wiki-knowledge-base fetch video-research -- https://docs.livekit.io/agents # fetch a specific page via Chrome into raw/
/wiki-knowledge-base fetch video-research -- https://docs.livekit.io       # fetch a docs root: grabs the 3-5 key sub-pages
/wiki-knowledge-base ingest video-research                                  # process new raw sources into wiki pages
/wiki-knowledge-base query video-research -- compare vendor latency         # search wiki, synthesize an answer
/wiki-knowledge-base lint video-research                                    # health-check the wiki
/wiki-knowledge-base status                                                 # list all wikis with stats

Arguments & flags

Command	What it does
`init <topic>`	Scaffold a new wiki (asks for the wiki’s focus)
`fetch <topic> -- <url> [url2] ...`	Fetch web pages via the Chrome browser extension into `raw/` — bypasses bot protection that blocks plain fetching. Point it at a docs root URL (rather than a specific page) and it navigates the nav/sidebar to fetch the 3-5 most important sub-pages instead of a single page
`ingest <topic>`	Process unprocessed `raw/` sources into interlinked wiki pages
`query <topic> -- <question>`	Search the wiki and synthesize a cited answer
`lint <topic>`	Find dead links, orphan pages, contradictions, stale sources
`status [topic]`	Dashboard of all wikis, or detail for one

What it does

Each wiki has three layers under .claude/wikis/<topic>/: immutable raw/ sources, the LLM-maintained wiki/ (index, log, pages), and a schema.md of conventions.

init — scaffolds the directory structure, then asks you to describe the wiki’s focus (user-confirmation checkpoint) and generates the schema, index, and log.
fetch — opens each URL in a Chrome tab, extracts the page text, and saves it as a markdown source in raw/. If a URL is a docs root rather than a specific page, it identifies and fetches the 3-5 most important sub-pages instead of crawling the whole site. Falls back to direct fetching if Chrome tools aren’t connected.
ingest — finds raw sources not yet in the log, reads each one, and presents 3-5 key takeaways for you to confirm or correct (user-confirmation checkpoint) before writing summary, entity, and concept pages, updating cross-references, the index, and the log.
query — reads the index, greps for relevant pages, reads only the matches, and synthesizes a cited answer. If the answer is novel synthesis, it offers to file it back as a new wiki page (user-confirmation checkpoint).
lint — checks for orphan pages, dead links, contradictions, index drift, stale sources, and missing cross-references; reports issues by severity with suggested next steps.
status — shows page/source counts, recent activity, and a page breakdown per wiki.

Agents spawned

Agent	Model tier	Role
explore-light	haiku	`lint` scan for large wikis (> 20 pages) — keeps lint cost flat

All other operations run inline in your session.

Output & artifacts

.claude/wikis/<topic>/raw/ — your immutable source documents (the LLM never modifies these)
.claude/wikis/<topic>/wiki/index.md — content catalog; wiki/log.md — append-only operation record; wiki/pages/ — entity, concept, summary, comparison, and synthesis pages
.claude/wikis/<topic>/schema.md — conventions, co-evolved with you
Other skills discover wikis automatically: /planning, /fix, /implement, and /review-pr grep the index and read only matching pages

Troubleshooting

Problem	Fix
`No new sources found` on ingest	Add documents to `.claude/wikis/<topic>/raw/` first (markdown, text, PDF)
`fetch` can’t reach a page (Chrome MCP not connected)	It falls back to direct fetching; if that 403s, save the page with Obsidian Web Clipper and drop the file into `raw/`
Lint reports dead links or index drift	Re-run `ingest` or fix the flagged pages — HIGH-severity issues list the suggested fix
Wiki getting slow to search (100+ pages)	The skill’s index-first protocol keeps reads selective; consider an external local-search tool for very large wikis

/kb-architecture-diagram — diagram the project knowledge base (a different KB than topic wikis)
/planning, /fix, /implement — consume wiki context automatically
/calibrate — the project knowledge base; wikis research, the project KB applies