Research Skill
Three modes. Select by request signal:
| Signal | Mode |
|---|---|
| Formal research, investigation, sourced report, gather evidence | Research Pipeline |
| Fact check, verify claims, check facts, is this accurate, verify quote | Fact-Check |
| Why did you, explain routing, show trace, decision log, why that agent | Explanation Traces |
Default: Research Pipeline.
Final report. When the final report synthesizes two or more dispatched agents' findings, run one self-review pass on the draft before delivery: skills/shared-patterns/report-self-review.md (separate dispatch, model: opus). It applies to Mode A and Mode B.
Mode A: Research Pipeline
Mode B: Fact-Check
Verify every factual claim in a draft before publish. Burden of proof sits on the claim, not the checker. Works standalone or as a pre-publish gate. Non-blocking: the report warns; the caller decides whether to publish.
Phase 1: EXTRACT
List every checkable claim: statistics, prices, dates, quotes, attributions, titles, event facts, rankings, causal assertions. Opinions and speculation stay out.
For each claim, record: ID, verbatim text, type (stat/quote/attribution/event/title/price/causal), location. Extract quotes verbatim for exact-words comparison.
Gate: Every checkable assertion has a claim ID. Sweep the document twice.
Phase 2: VERIFY
Work claim by claim.
- Check provided sources first. Search caller-supplied documents before external sources. Record exact passages.
- Read laterally. Judge a source by what other sources say about it, not by its own presentation.
- Climb the source tier. Follow citations upward: primary (study, filing, transcript) > direct secondary (interviews, primary reading) > derived (aggregators, rewrites). A broken citation chain caps the claim at Unverifiable.
- Triangulate contested claims. Two independent sources (separate origins, not wire rewrites). One source suffices for routine facts from a primary document.
- Verify quotes on three axes. All must hold: exact words match, attributed speaker confirmed, original context supports the meaning used.
- Check staleness. Time-sensitive claims expire. Prices/rates: 1 day-1 week. Counts: 1-3 months. Titles/roles: 3-6 months. Event status: until event date. Surveys: 6-12 months. Science: 1-3 years. Laws: 6-12 months. Records/superlatives: re-check every use. Stable history: none.
Gate: Every claim has an evidence record.
Phase 3: ADJUDICATE
Assign each claim one label:
| Label | Assign when |
|---|---|
| Verified | Evidence supports; current within staleness window; sufficient source tier |
| Disputed | Evidence contradicts; newer source supersedes; quote fails any axis |
| Unverifiable | Sources engage the claim but settle nothing |
| Missing-source | No available source addresses it |
Rules: contradiction beats support. Partial verification gets the weakest label. Stale figure superseded by newer = Disputed; merely old with no newer figure = Unverifiable.
Gate: Every claim carries one label and a one-line justification.
Phase 4: REPORT
# Fact-Check Report: [document] ## Summary Claims: N | Verified: n | Disputed: n | Unverifiable: n | Missing-source: n Unchecked: n (reason) ## Per-Claim Findings ### C1 -- [label] Claim: [text] | Evidence: [source + passage] | Reasoning: [why this label] ## Warnings [Every Disputed/Missing-source claim with correction] ## Publish Recommendation [Hold / fix-then-publish / clear]
Every time-sensitive Verified claim carries its as-of date.
Gate: Report covers every claim ID. Warnings lists every Disputed and Missing-source finding.
Mode C: Explanation Traces
Read the per-dispatch route event log and present routing decisions as a human-readable timeline. Answer "why did I get routed here?" from recorded events only -- never from reconstruction or rationalization.
Log path: ${CLAUDE_LEARNING_DIR:-$HOME/.claude/learning}/route-events.jsonl (append-only JSONL).
Phase 1: LOCATE
bashLOG="${CLAUDE_LEARNING_DIR:-$HOME/.claude/learning}/route-events.jsonl" wc -l "$LOG"
If absent or empty, report the path and the producing hook (hooks/routing-decision-recorder.py). Do not reconstruct from memory.
Phase 2: PARSE
Parse each JSON line. Two event types: DECISION (one per /do-routed dispatch) and OUTCOME (one per finalized dispatch). See references/trace-schema.md for full field semantics.
Filter to the user's query:
| User signal | Filter |
|---|---|
| Names an agent or skill | DECISION/OUTCOME events matching that name |
| "Why routed here" / latest | Most recent DECISION, current session first |
| Outcome question | OUTCOME events, joined to decisions |
| No specific target | Chronological timeline, most recent session |
Join OUTCOME to DECISION on same session AND key == "{agent}:{skill}". File adjacency is unreliable.
Phase 3: PRESENT
Sort by ts. For each decision, show: time, agent+skill, complexity, request snippet, health at decision (three states: numeric, no-weight-row, legacy), alternates, outcome.
Lead with the answer to the user's specific question, then offer surrounding context. Flag gaps honestly: pre-instrumentation entries, unmatched outcomes.
request_snippet is private session data: show to the session's user, keep out of PR bodies, issues, exports.
Deep References
| Signal | Reference | Content |
|---|---|---|
| Event field schema, health states, join rules | references/trace-schema.md | DECISION/OUTCOME field semantics |
| Diagnosing thin trace data, consumer mistakes | references/preferred-patterns.md | Failure mode catalog for log reading |
| Parse/read errors, missing log, unmatched outcomes | references/error-handling.md | Error-fix mappings for trace reading |
