| name | researching |
| description | Use when you need to understand how an existing codebase works and/or survey external prior work (papers, methods, tools) before planning or changing something. Traces function call paths, maps module dependencies, searches academic papers and documentation, compares library alternatives. Triggers: research the codebase, how does X work, where is X implemented, what tools or libraries exist for X, prior art on X, has this been done before, survey approaches to X. |
Researching
Build the context a task needs by investigating the current codebase and/or
external prior art in one pass, then synthesizing a single research document
that feeds planning. Run the internal dimension, the external dimension, or
both.
Interaction mode
Leans Collaborative by default; see ai-research-workflows:using-research-workflows for the full protocol.
Checklist
Work through these in order and track them as tasks. Do not skip the review gates (steps 5, 7, and 8):
- Explore context — the request, existing
docs/rse/specs/ (and legacy .agents/) docs, recent commits
- Scope — internal / external / both; what counts as relevant; depth
- Propose the research plan — sub-questions + sources; get a quick OK
- Investigate — run the codebase and/or prior-art pass
- Present findings — in sections; confirm understanding before writing anything out
- Write the research document —
docs/rse/specs/research-<slug>.md from the template
- Self-review — read the written doc with fresh eyes; fix issues inline
- User reviews the document — have them go through the written doc before you move on
- Hand off — point to the next skill
Process
Scope → plan → investigate → present → write → self-review → user review → hand off.
1. Explore context
Read the request. Scan docs/rse/specs/ (then legacy .agents/) for existing research and plan documents
(research-*.md, and legacy prior-art-*.md). Note recent commits if relevant.
2. Scope the question
Establish: the precise question; whether it needs internal codebase research,
external prior-art research, or both; what counts as relevant; and the
desired depth. In Collaborative mode ask one focused question at a time; in
Direct mode infer scope from the request and state it before proceeding.
3. Propose the research plan
Before investigating, present the decomposition: the codebase sub-questions you
will trace and/or the prior-art sources and dimensions you will search. Get a
quick OK (Collaborative) or state it and proceed (Direct).
4. Investigate
Run the dimension(s) in scope:
- Codebase pass — see
references/codebase-research.md
- Prior-art pass — see
references/prior-art-research.md
Both passes may surface gaps and light, clearly-labeled observations; defer deep
critique and design decisions to planning or validation.
5. Present findings
Present synthesized findings in sections (codebase first, then prior art),
confirming understanding as you go. Use file:line anchors for code and
citations for external sources.
6. Write the research document
Derive a slug (lowercase, hyphenated). Read the template:
assets/research-template.md. Fill the sections that apply and write to
docs/rse/specs/research-<slug>.md. Omit sections for any dimension not investigated.
If the internal pass ran, record the codebase state you researched against —
the commit SHA (git rev-parse --short HEAD) and the date — so file:line
findings can be re-checked as the code evolves.
7. Self-review
Before showing the document, read it through with fresh eyes and fix issues inline:
- Placeholders — no "TBD"/"TODO", empty sections, or vague findings.
- Anchors & citations — every codebase finding cites a
file:line; every prior-art claim cites a source.
- Consistency & coverage — sections don't contradict, the synthesis matches the findings, and every scoped sub-question is answered (or its gap is named).
Confirm the doc meets the Quality checklist below.
8. User reviews the document
Surface the written document and let the user go through it before you move on:
Research doc written to docs/rse/specs/research-<slug>.md. Please review it and tell me if anything is missing or off before we move on to planning.
In Collaborative mode, wait for their response; if they request changes, revise and re-run the self-review. In Direct mode, present the doc and proceed, noting it is ready for their review.
9. Hand off
Point to the next skill — usually ai-research-workflows:planning-implementations
(or ai-research-workflows:running-experiments when an approach is uncertain).
Common Mistakes
- Researching both dimensions when one suffices — scope first; don't search external prior art for a question that is purely about current code, or vice-versa.
- Uncited prior-art claims — every external factual claim needs a primary-source citation; label secondary sources as such.
- Vague codebase findings — every code finding cites a specific
file:line; "it's somewhere in auth" is not a finding.
- Skipping the disconfirming search — for prior art, actively search for limitations, critiques, and failures, not just supporting evidence.
- Turning research into a rewrite plan — surface gaps and light observations, but defer detailed design and critique to planning/validation.
Quality checklist
Cross-references
- Feeds
ai-research-workflows:planning-implementations.
- For comparing uncertain approaches with real code, use
ai-research-workflows:running-experiments.
- Surveyed automatically by
ai-research-workflows:using-research-workflows.