bjornmelin/context7-research

Context7 Research

Use Context7 as the source of truth for library/framework documentation during implementation, reviews, and debugging.

Workflow (MUST FOLLOW)

1. Extract:
   • libraryName: product/package/framework name (e.g., next.js, zod, supabase)
   • objective: what you need to do (e.g., “configure SSR auth”, “migrate v3->v4”, “use hook X”)
   • version: only if the user specifies one or you can infer it from repo context (otherwise default to “latest”).
2. If the user already provided a Context7 library ID in /org/project (or /org/project/version) format, skip resolution.
3. Otherwise call mcp__context7__resolve-library-id with:
   • libraryName: extracted name
   • query: the full user objective (not just a keyword)
4. Select the best match using this rubric:
   • Exact/closest name match
   • Highest source reputation
   • Highest snippet coverage (prefer more snippets when reputation is similar)
   • Highest benchmark score
   • Versions available that match the requested version (if any)
5. Call mcp__context7__query-docs with the selected libraryId and a tight query (see Query Budget).
6. Synthesize an answer:
   • Include the libraryId used (and version if applicable)
   • Quote or paraphrase only what’s needed; prefer code examples
   • If docs are missing/ambiguous, label the gap as UNVERIFIED and ask for the smallest clarification needed.

Query Budget (HARD LIMIT)

Use at most 3 mcp__context7__query-docs calls per user request. Plan the queries up-front:

• Query 1 (overview): “Explain the concept and where it lives in the docs; include minimal example.”
• Query 2 (API details): “Exact function/class/method signatures and usage examples for the user’s task.”
• Query 3 (edge cases): “Pitfalls, errors, version differences, and recommended patterns.”

If you can’t cover everything within 3 calls, ask the user to narrow scope (or pick the highest-impact subquestion).

Output (DEFAULT SHAPE)

• Start with the recommended approach (1–5 bullets).
• Include at least one doc-backed example/snippet when the question is about code.
• End with “Assumptions / Gaps” if anything is uncertain (UNVERIFIED) or version-dependent.

Resources

• Playbook: references/playbook.md
• Troubleshooting: references/troubleshooting.md
• Design + scoring (maintainers): references/design-spec.md
• Report template: assets/report-template.md
• Generate a new report file: python3 scripts/new_report.py --out ./context7-research.md --library-name "<name>" --question "<objective>"