research-web ยท diff
git:20260924.17234b5 to git:20261105.60e9622
17 added, 8 removed. Audit A to A.
---
name: research-web
- description: Perform technical web research using authoritative sources, synthesize the results, and save a cited research note in the repo. Use when external documentation or up-to-date behavior matters.
+ description: Perform technical web research using authoritative sources, synthesize the results, and save a cited research note in the repo. Use when external documentation or up-to-date behavior matters, or when handed one question the codebase cannot answer.
---
# Research Web
- Produce a research note with citations and direct links.
+ Produce a cited answer with direct links.
Workflow:
- 1. Break the topic into 2-5 research dimensions.
- 2. Prioritize official documentation, primary sources, and recent material.
- 3. Browse broadly first, then open only the most promising sources.
- 4. Compare sources, note conflicts, and verify version details.
- 5. Save the result to `./tmp/research/YYYY-MM-DD-web-description.md`.
- 6. Present the summary, confidence level, and open questions.
+ 1. Restate the question and what would count as an answer. Break it into 2-5 research dimensions.
+ 2. Prioritize official documentation, primary sources, and recent material, and check it applies to the version in use.
+ 3. Browse broadly first, then open only the most promising sources. Cross-check anything that matters against a second source.
+ 4. Compare sources, note conflicts and which is more authoritative, and verify version details.
+ 5. Save the result to `./tmp/research/YYYY-MM-DD-web-description.md`. When another agent asked the question, return the report to it instead.
+ 6. Present the report.
+ Report:
+ - The answer, in a few sentences
+ - Findings: claim, source, date or version
+ - Conflicts, gaps, and open questions
+ - Confidence: high, medium, or low
+
Requirements:
- Every factual claim needs a source URL.
- Distinguish official guidance from community opinion.
- Call out software versions, dates, and deprecations.
+ - Say plainly what you could not verify. Never fill a gap with a plausible guess.
+ - Stay inside the time box if you were given one.
- Prefer concise synthesis over long quote dumps.
+ - When answering for another agent, report findings only; it decides the design.