Wiki Query

Parameters

Workflow

1. Locate the wiki and read the index

Walk upward to find .cogni-wiki/config.json. Read wiki/index.md top to bottom. The index is the map — do not skip it.

2. Select candidate pages

From the index, select pages whose one-line summary is relevant to the question. If fewer than 2 clear matches, also:

• Run grep -l -i over wiki/pages/ for the question's key nouns
• Cap the total set at --max-pages

If zero candidate pages emerge after both passes, stop and report: "The wiki has no pages on this topic. Run wiki-ingest on a source first, or rephrase the question." Do not answer from memory.

3. Read the selected pages

Read each candidate page fully. Note:

• The claims that directly answer the question
• The sources each claim traces to
• Any contradictions between pages (flag for wiki-lint if you find one)

4. Synthesize the answer

Write a plain-prose answer, inline-citing every factual claim with a [[page-slug]] link. Example:

Constitutional AI replaces human harm labels with AI-generated critiques against a written constitution [[constitutional-ai]]. The method scales with critic-model size [[constitutional-ai]] and was introduced by Bai et al. in 2022 [[anthropic-safety-team]].

Rules:

• Every non-trivial claim links to at least one [[page]]
• If two pages contradict each other, surface the contradiction explicitly ("[[page-a]] claims X, but [[page-b]] claims Y — these should be reconciled via wiki-update or flagged via wiki-lint")
• If the wiki's coverage is thin, say so: "Based on two pages, both summarizing one 2022 paper, the wiki's view is..."

5. Offer to file the answer back

If --file-back auto, ask the user: "File this answer as a new wiki page?" If yes, pass control to the file-back step below. If --file-back yes, do it without asking. If no, stop here after step 4.

6. File-back (optional)

If filing the answer back:

1. Choose a slug derived from the question or the synthesized claim
2. Write a new page at wiki/pages/{slug}.md with type: learning (or summary if the answer is mostly one source)
3. Frontmatter sources: field lists the wiki pages used, prefixed with wiki:// to distinguish from raw sources:

   sources:
     - wiki://constitutional-ai
     - wiki://anthropic-safety-team
   

   This is explicit: the learning is a derived synthesis, not a direct restatement of a raw source.
4. Body is the synthesized answer with the [[citations]] preserved
5. Update wiki/index.md under an appropriate category
6. Append to wiki/log.md:

   ## [YYYY-MM-DD] query | {slug} — {short question}
   

7. Increment entries_count in .cogni-wiki/config.json
8. (Optional) Run backlink_audit.py from wiki-ingest on the new page to add bidirectional links

7. Always append to the log (even without filing back)

Regardless of file-back decision, append a query log line:

## [YYYY-MM-DD] query | "{short question}" → read {N} pages

The log records every query so the user can see what the wiki has been asked.

Output

• An answer with inline [[citations]]
• Optionally: a new wiki/pages/{slug}.md learning page
• An appended line in wiki/log.md (always)
• Optionally: updated index, config, and backlinks (if file-back was yes)

Golden rules

1. Answer only from wiki content. The user chose a wiki precisely because they want answers grounded in curated sources, not in training data that may be stale or wrong. If the wiki does not contain a claim, the answer does not contain that claim.
2. Every factual sentence is [[cited]]. Rhetoric and connective text can go uncited; claims cannot.
3. Contradictions are surfaced, not reconciled. Only wiki-update resolves contradictions; wiki-query reports them.
4. Thin coverage is declared. If the wiki has one page on a topic, say so — don't pretend to a consensus.
5. Always log. Every query leaves a trail, filed-back or not.

References

• ${CLAUDE_PLUGIN_ROOT}/references/karpathy-pattern.md — the pattern
• ./references/query-patterns.md — read-before-answer worked example