| name | blcaptain-research-driven |
| description | Improves research quality when accuracy matters; work depends on external facts such as documentation, APIs, pricing, licenses, standards, current events, research, or source code; a question is complex or multi-part; confidence is low; or sources conflict. Excludes casual chat, pure formatting, mechanical edits, and trivial facts already known with high certainty. |
| license | Apache-2.0 |
Research-Driven Investigation
Investigate; do not merely retrieve.
Treat a plausible answer as the start of inquiry, not its end. Build a conclusion that survives attempts to disprove it. Maximize correctness without promising certainty. Expose uncertainty instead of filling gaps.
Protect the Two Highest-Priority Behaviors
Try to disprove before supporting
- State the strongest competing explanation or answer.
- Predict what evidence would appear if that alternative were true.
- Search for counterexamples, boundary failures, changed assumptions, and contrary primary sources.
- Revise or narrow the conclusion when contrary evidence survives scrutiny.
Do not call a search complete after collecting only supporting material.
Stop only after convergence
Stop when every applicable condition holds:
Continue when any applicable item remains false. Stop earlier for a trivial or low-impact task when one authoritative source directly answers the whole question and no meaningful conflict, ambiguity, or freshness risk exists. State that reduced scope.
Use Three Levers
| Lever | Make it operational | Avoid |
|---|
| Thorough | Decompose the question into mutually exclusive, collectively exhaustive sub-questions; track coverage; apply the explicit STOP conditions. | Collecting many links without proving coverage. |
| Deep | Prefer the closest primary source; inspect definitions, mechanisms, boundaries, version history, and counterexamples; ask why the fact holds. | Repeating summaries that merely agree. |
| Correct | Cross-verify material claims; test alternatives; distinguish source independence; grade certainty. | Treating consensus, rank, or confidence as proof. |
Run the Seven-Step Investigation
1. Frame and decompose
- Restate the exact question, intended decision, audience, time boundary, and required precision.
- Separate facts to establish from judgments to make.
- Split the question into MECE sub-questions.
- Add boundary questions for version, jurisdiction, date, environment, population, and excluded cases.
- Mark dependencies and what could change the answer.
- Calibrate effort to impact, uncertainty, reversibility, and freshness.
Use the coverage and overlap tests in references/techniques.md for complex decomposition.
2. Survey the landscape
- Map the terrain before narrowing the search.
- Identify authoritative source families, disputed areas, vocabulary, versions, and likely blind spots.
- Name the best available primary source and an independent corroborating source for each sub-question.
- Read references/tool-mapping.md before relying on platform-specific tool names.
- Keep the survey broad but brief; use it to plan depth, not to form the final conclusion.
3. Trace to primary sources
- Search one sub-question or hypothesis at a time.
- Open the underlying source; do not treat a search snippet as evidence.
- Follow citations upstream until reaching the closest authoritative origin available.
- Record the exact passage, table, behavior, date, version, and scope that support or weaken the claim.
- Ask why the observed fact holds; identify its mechanism, definition, assumptions, and failure boundary.
- Separate observed fact, source interpretation, and your inference.
For persistent work, copy assets/research-log-template.md and fill it while investigating.
4. Cross-verify
- Corroborate each key conclusion with at least two genuinely independent sources.
- Trace source ancestry; reject circular citation, mirrored content, or repeated claims from a shared origin as independent support.
- Compare sources that use different methods when possible.
- Resolve disagreements by checking scope, date, definitions, methods, incentives, and source ancestry.
- Downgrade the conclusion when independent corroboration remains unavailable.
5. Actively disprove
- Steelman the strongest alternative before rejecting it.
- Predict what evidence would appear if that alternative were true.
- Search using contrary phrasing and failure-oriented terms.
- Seek counterexamples, boundary cases, version differences, and the strongest opposing evidence.
- Downgrade, narrow, withdraw, or replace a conclusion when it does not survive the challenge.
6. Identify gaps and iterate
- Name every unresolved sub-question, conflict, access limit, and untested boundary.
- Reframe weak questions in light of new evidence.
- Loop through survey, primary-source tracing, cross-verification, and disconfirmation as findings demand.
- Track what each new round changes.
- Apply the STOP checklist; continue only when another round could materially change the result.
7. Synthesize with graded certainty
- Answer the decision first.
- Show the minimum reasoning needed to make the answer auditable.
- Attach sources to the claims they support.
- Grade each material conclusion as Confirmed, Probable, Speculative, or Unknown.
- State versions, dates, assumptions, exceptions, and open questions plainly.
- Reapply the STOP checklist before finalizing.
Grade Certainty
| Grade | Assign only when |
|---|
| Confirmed | At least two genuinely independent primary sources agree, material conflicts are resolved, and active disconfirmation found no surviving contradiction within the stated scope. |
| Probable | Credible evidence supports the claim, but full independent primary-source corroboration or one meaningful boundary test remains incomplete. |
| Speculative | Only reasoning or weak, indirect evidence supports the claim, and plausible alternatives remain open. |
| Unknown | Available evidence cannot determine the answer; keep the gap open instead of guessing. |
Do not convert source confidence, model confidence, or repeated wording into a higher grade. State scope beside the grade when a conclusion depends on a version, date, environment, or population.
Present the Result
Use this compact structure unless the user requests another format:
- Conclusion: State the decision-relevant answer and certainty grade.
- Why: Give the decisive facts and reasoning.
- Countercheck: State the strongest alternative tested and what happened.
- Boundaries: Name versions, dates, assumptions, exceptions, and unresolved gaps.
- Sources: Link each material claim to the closest supporting source.
Avoid Research Failure Modes
- Shallow stop: Do not stop at the first plausible answer.
- Single-source conclusion: Do not treat one source as corroboration.
- Confirmation bias: Do not gather only supporting material; hunt for the strongest contrary evidence.
- Secondary as authority: Do not prefer commentary when an accessible primary source can answer the question.
- Circular corroboration: Do not count pages that repeat one upstream claim as independent confirmation.
- Speculation as fact: Do not hide an inference, unresolved conflict, or guess behind confident prose.
- Boil the ocean: Do not overresearch low-impact questions after the STOP conditions hold.
- Do not claim exhaustive coverage without a decomposition and a coverage check.
- Do not confuse "no contradiction found" with proof that none exists.
- Do not guarantee correctness; make the remaining uncertainty visible.
Load Supporting Material Only When Needed
