wiki-auto-research

name: wiki-auto-research description: Scan wiki/ for Open Questions and ambiguous claims; web-search gaps; write reviewable notes to research/ with generated_by marker. Manual one-shot. Do NOT use for general web research (use WebSearch). Obsidian wiki 知識穴埋め・知識補完・自動調査。

Wiki Auto-Research — Fill Knowledge Gaps via Web Search

Scans the wiki for unanswered questions and low-confidence claims, runs targeted web searches, and writes structured research notes to research/ for user review. Notes are NOT automatically merged — the user reviews and runs /wiki-ingest to integrate.

See research-note-format.md for the output contract.

Trigger model

[!important] This skill is manual one-shot. It does NOT auto-fire. It does NOT loop. It does NOT chain into /wiki-ingest. Each run is bounded and produces reviewable artifacts.

Pre-flight

1. Read .obsidian-wiki.config for OBSIDIAN_WIKI_VAULT_PATH (defaults wiki/). If only legacy .env exists, instruct user to run /wiki-setup to migrate.
2. Check WebSearch tool availability — this skill is useless without it. Verify by:
   • Inspecting the tool list in this session (look for WebSearch in the available tools)
   • If absent, exit with this message:

     wiki-auto-research requires the WebSearch tool, which is not available in this session.
     
     Likely cause: this skill is running in a sandboxed environment (e.g., Cowork)
     where outbound URL fetches are blocked.
     
     Run this skill from Claude Code CLI, or use a different approach:
     - Hand-research the question and update wiki/ pages directly
     - /wiki-ingest after you write a manual research note
     

3. Confirm <vault-root>/research/ exists; create with mkdir -p if missing.

STEP 1 — Identify candidate gaps

Scan all wiki pages for:

Open Questions

Body sections under ## Open Questions. Each bullet is one candidate question.

Low-confidence claims

• ## Key Facts bullets ending in ^[ambiguous] — these flag where sources disagree
• ## Summary ending in Confidence: unverified — page-level low confidence
• frontmatter.status: seed pages older than 30 days — likely stalled at seed

Contradictions

Pages with ## Contradictions sections — research can adjudicate.

STEP 2 — Filter by idempotence

For each candidate question:

1. Read all existing research/*-auto-research.md notes
2. Collect their source_questions: frontmatter values (any status, including pending-review)
3. Skip any candidate question that's already in that set

Unless user passes --force <topic> to override.

This prevents re-researching already-covered ground.

STEP 3 — Show plan, get user OK

Found 12 unique gaps:

PRIORITY (with explicit Open Questions):
  1. entities/Thompson-Sampling — "Are there 2025 empirical benchmarks?"
  2. concepts/exploration-exploitation — "How does context-free MAB compare to contextual?"
  3. ...

LOW-CONFIDENCE (ambiguous facts):
  4. entities/qlib — Key Fact about supported brokers is ^[ambiguous]
  ...

CONTRADICTION:
  9. synthesis/finance-ai — "Source A says X; Source B says Y"

I'll batch related questions into research notes (likely 6–8 notes total).
Estimated web searches: ~25 (3–5 per topic cluster).

Proceed? (yes / select-specific / skip)

Wait for user response. Respect "select-specific" by asking which to research.

STEP 4 — Cluster & batch

Group related questions into topic clusters per research-note-format.md "Per-question vs. per-topic batching" rules:

• Same source page → same cluster
• Same domain + thread → same cluster
• Single set of web searches answers them → same cluster

Each cluster becomes one research note.

STEP 5 — Per-cluster research loop

For each cluster:

5a. Plan searches

Draft 3–5 targeted web search queries. Vary phrasing to surface different angles. Record queries for the note's search_queries_used: field.

5b. Execute searches

Use WebSearch. Record URLs, titles, key passages.

5c. Synthesize

Per research-note-format.md:

• Lead with bottom-line answer
• Cite sources inline
• Note confidence + disagreements
• Recommend specific wiki updates

5d. Write the note

Output to <vault-root>/research/YYYY-MM-DD <topic-slug>-auto-research.md with full frontmatter (generated_by: wiki-auto-research, status: pending-review).

5e. Pause if scope is large

After every ~3 notes, give the user a progress update and option to pause:

Done with 3 of 8 clusters. Continue? (continue / pause)

This prevents long unchecked runs.

STEP 6 — Append to log and report

Append to wiki/log.md:

| YYYY-MM-DD | auto-research | <N> questions across <M> clusters | research/<filename1>, research/<filename2>, ... |

Final report:

Auto-research complete:
  Questions covered:  12
  Clusters → notes:    8
  Web searches:       29

Output:
  research/2026-05-03 thompson-sampling-2025-empirical-auto-research.md
  research/2026-05-03 mab-context-comparison-auto-research.md
  ...

Next steps:
  1. Review each note (set status to reviewed-accept or reviewed-reject)
  2. Run `/wiki-ingest` with scope "Research notes only" to merge accepted notes
  3. Or hand-edit notes first, then ingest

Notes are NOT merged into wiki/ until you run /wiki-ingest.

Boundaries (what wiki-auto-research does NOT do)

• ❌ Does not modify any wiki page directly (output is research/, requires user review + /wiki-ingest)
• ❌ Does not loop / auto-trigger / chain into other skills
• ❌ Does not re-research already-covered questions (idempotence via source_questions: set)
• ❌ Does not exceed ~5 searches per cluster without user OK
• ❌ Does not fabricate sources — every claim cites a URL or is marked low-confidence
• ❌ Does not delete Open Questions from wiki pages (that's /wiki-ingest's job after user accepts)

Failure modes to watch

Rate limit / cost awareness

Cap per run: ~30 web searches by default. If estimated count exceeds, batch and ask user to confirm before continuing.