Skip to content
Back to skills

Research

ASecurity

Investigate a technical question against primary sources and capture the findings, with citations, as a file in the repo. Use when the user wants a topic researched, docs or API facts gathered, a claim verified before it's relied on, or reading legwork delegated to a background agent.

  • 18 stars
  • 0 votes
  • 0 copies
  • 3 views
  • Added August 31, 2026
developmentgoreactapidocumentation

Works with

  • api

Security analysis

A100/100

Pro scans all 2 files and shows the line behind each finding

Scanned August 31, 2026

npx -y skills add aethrox/doctrine --skill research --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Research?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Research
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/aethrox-research/badge)](https://www.skillsdirectory.com/skills/aethrox-research)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: research
description: Investigate a technical question against primary sources and capture the findings, with citations, as a file in the repo. Use when the user wants a topic researched, docs or API facts gathered, a claim verified before it's relied on, or reading legwork delegated to a background agent.
---

# Research

A claim is only as good as the source it traces back to. This skill's defining constraint: every finding in the output must be traceable to the primary source that actually owns the fact (official docs, source code, a spec, a first-party API response), not to a blog post, forum answer, or summary that itself read the primary source once. If tracing stops at a secondary source, that's stated as a limitation, not hidden.

## Phase 1: Scope the question

State the specific question or claim being investigated, narrowly enough that "found" and "not found" are both clear outcomes. A vague scope ("research React") produces an unfalsifiable report; a narrow one ("does React's `useEffect` cleanup run before or after the next effect's setup, per the official docs") produces a checkable answer. If the request bundles several questions, split it, mixing them is how one weak answer contaminates the credibility of the rest.

## Phase 2: Trace every claim to its primary source

For each sub-question:

1. **Go to the source that owns the fact.** Official documentation, the actual source code, a published spec or RFC, a first-party API response, not a secondary write-up of any of these.
2. **If only a secondary source turns up, follow its own citation back.** A blog post that says "per the docs, X" is a pointer, not the finding; open what it points to and verify the docs actually say X before citing the docs, not the post.
3. **If tracing dead-ends at a secondary source** (the primary source is gone, paywalled, or the secondary source cites nothing verifiable), say so explicitly in the finding rather than presenting a secondary source as if it were primary.
4. **Note the primary source's own date or version.** A fact that was true in an old version of a doc or API may not be true now; carry the version/date forward into the finding, not just the claim.

## Phase 3: Weigh source credibility

Where more than one primary-adjacent source disagrees, or where authority isn't obvious, check the source against the CRAAP criteria (Currency, Relevance, Authority, Accuracy, Purpose; the standard library-science checklist for source evaluation):

| Criterion | Question |
|---|---|
| **Currency** | Is this the current version, or superseded by a newer release/spec? |
| **Relevance** | Does it actually answer the scoped question, or just something adjacent? |
| **Authority** | Is the author/publisher the party that actually owns this fact (the framework's own maintainers, the spec body) or a third party writing about it? |
| **Accuracy** | Can the claim be cross-checked against another independent primary source, or the actual behavior (running the code, reading the actual response)? |
| **Purpose** | Is the source explaining the fact, or selling/promoting something that gives it a reason to shade the answer? |

A source that fails Authority or Purpose isn't disqualified outright, but the finding should say so; "per a third-party benchmark, not the vendor's own numbers" is a materially different claim than "per the vendor's published spec."

## Phase 4: Write the findings, cited

Capture the investigation as a single Markdown file:

- Each finding states the claim, then the source it traces to (link or exact reference; file path, doc URL, spec section), and the source's date/version where that matters.
- Conflicting sources are surfaced side by side, not silently resolved by picking one; the reader deciding what to do with the conflict is a different judgment call than the researcher's.
- Anything that couldn't be traced past a secondary source is marked as such, not smoothed over.
- Save it where the repo already keeps this kind of note; if there's no existing convention, pick a sensible location and say where, so the next research pass doesn't have to rediscover it.
- Write the report without em dashes: use a comma, colon, period, or a reworded sentence instead.

Delegate the legwork to a background agent when the investigation is multi-step or time-consuming, so the primary conversation isn't blocked on reading, but the citation discipline in Phases 2–4 applies regardless of who does the reading.

## Done when

- [ ] The question is scoped narrowly enough to have a checkable answer.
- [ ] Every finding traces to a primary source, or explicitly states that it doesn't.
- [ ] Disputed or authority-weak sources are flagged per the CRAAP criteria rather than presented as settled.
- [ ] The output is a single cited file, saved where the repo's existing convention expects it (or a stated new location).

Files in this skill

  • SKILL.md4.8 KB
  • agents/openai.yaml127 B

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.