Research Way¶
Source: hooks/ways/research/research.md
Frontmatter
| Field | Value |
|---|---|
description |
Structured investigation — scoping a question, fanning out across independent sources, ranking source authority from official docs down to community posts, and synthesizing findings into an answer |
vocabulary |
research investigate look into dig into find out compare evaluate assess synthesize source sources primary source official docs authority credible trust citation evidence survey landscape alternatives options fan-out sweep confidence tangent comparative recommendation |
refire |
0.15 |
macro |
append |
scope |
agent |
requires |
['Bash(grep:*)'] |
Define the Question¶
Before searching, state what you're trying to learn and why. A vague "research X" becomes "I need to understand X to decide between Y and Z." The user may not have articulated this — help them sharpen it.
Investigation Structure¶
- Scope — What's in bounds? What would be a tangent?
- Gather — Use tools (WebSearch, WebFetch, Grep, Read) to collect information. Prefer primary sources over summaries of summaries. When the question spans several independent sources, fan
Exploreorgeneral-purposeagents across them if the user asked for breadth, a skill in use directs it, or a CLAUDE.md line covers it, and say why in a clause ("six independent sources; agents return conclusions, not page dumps"). Otherwise, where your instructions restrict the Agent tool, name the fan-out and its width in one line and wait. Deep-research andWorkfloware proposals to discuss. See meta/subagents. Put four rules in every fan-out brief: - Say "not found" rather than guess.
- Every claim carries a file path or URL and an exact value (version, port, name).
- Sample rather than bulk-read: name the manifests, entry points, and config files to start from.
- End with what was deliberately omitted, so you can ask for it.
- Evaluate — Not all sources are equal. Rank them: official docs for the exact version in use, then upstream source, then migration guides and advisories, then community posts, then LLM-generated content. Flag confidence levels.
- Synthesize — Compress findings into a structure the user can act on. Don't dump raw results.
- Present — Lead with the answer, then the evidence. The user wants the conclusion first.
Comparative Analysis¶
When comparing options (tools, approaches, vendors, libraries):
| Column | Purpose |
|---|---|
| Option | What's being compared |
| Strengths | What it does well |
| Weaknesses | Where it falls short |
| Fits when | The scenario where this option wins |
| Avoid when | The scenario where it's wrong |
End with a recommendation. Don't just present a neutral table — defend a choice based on the user's context.
Source Discipline¶
- Distinguish what you know (training data) from what you found (tool results) from what you inferred.
- When citing web results, include the URL so the user can verify.
- If you can't find reliable information, say so. "I couldn't find authoritative sources on X" is more useful than hedging.
- Stale information is worse than no information. Check dates on sources.
When Research is Overkill¶
Not every question needs a structured investigation. If the answer is in the codebase, read the code. If it's a factual recall question, just answer it. This way is for genuine unknowns that require gathering and synthesis.
See Also¶
- architecture/design(softwaredev) — research feeds design decisions
- think(meta) — structured thinking strategies for investigation
- research/sweep(research) — child: a survey across many independent sources
- research/provenance(research) — child: every claim, figure and citation traced to its source