research

Structured multi-source research protocol with data freshness validation — use when gathering factual data from the web, GitHub, npm, or docs to produce sourced, timestamped, cross-referenced findings

SKILL.md

GitHub

Research Skill — chainskills

Protocol

Step 1 — Clarify scope

Define the research questions before searching. Map each question to a decision it informs.

Step 2 — Memory check

Query conversation history for prior research on overlapping topics. Reuse FRESH findings (< 90 days). Skip Step 4 for already-verified facts.

Step 3 — Workspace scan

Read: package.json deps, ROADMAP, AGENTS.md, relevant source files, existing templates. Note current state before looking externally.

Step 4 — External research (parallel where possible)

For each question, query >=2 independent sources from the authority hierarchy:

  1. Official registry (npm, crates.io, PyPI)
  2. Official docs / spec / RFC
  3. Official GitHub repo (README, releases, CHANGELOG)
  4. Verified community resource (MDN, caniuse, awesome-*)
  5. Tech blog / article
  6. Forum / StackOverflow

AI memory alone is never sufficient — always fetch from sources 1-4.

Step 5 — Cross-reference

For critical facts (versions, breaking changes, security), verify with >=2 independent sources. Flag discrepancies: DISCREPANCY: source A says X, source B says Y.

Step 6 — Freshness stamp

Assign a status to every external claim:

  • FRESH — retrieved < 90 days ago
  • AGING — 90 days to 1 year
  • STALE — 1 to 2 years
  • EXPIRED — > 2 years
  • UNVERIFIED — not fetched, AI memory only

Step 7 — Structured output

Produce a report with:

  • Source table (claim / source / URL / date / freshness / confidence H|M|L)
  • Dependency audit table (package / pinned / latest / gap / advisory)
  • Workspace vs external delta (what the codebase assumes vs what is actually true)
  • Stale/unverified warnings
  • Recommended next steps

Anti-patterns

  • Do not cite a URL without accessing it — always fetch
  • Do not treat AI training data as a source — always verify
  • Do not draw single-source conclusions for critical decisions
  • Do not omit retrieval dates
  • Do not start external research before checking workspace state

Output template

### Research Report — {topic} — {YYYY-MM-DD}

#### Workspace Findings
{What the codebase already knows/does about this topic}

#### External Findings
| Claim | Source | URL | Date | Freshness | Confidence |
|-------|--------|-----|------|-----------|------------|

#### Dependency Audit
| Package | Pinned | Latest | Gap | Advisory |

#### Stale / Unverified
{Items to re-verify before acting}

#### Recommended Next Steps
{Handoff-ready actions}