| model | opus |
| name | my-research |
| description | Deep codebase research with verified findings. Uses focused discovery agents, cross-checks claims against code, saves a durable research artifact, and records assumptions. |
Research Codebase
Answer the research question with verified, code-grounded findings. External docs, tickets, and notes are starting points, not substitutes for reading code.
Load Rules
Read these first:
~/.claude/rules/question-policy.md~/.claude/rules/context-checkpoint.md~/.claude/rules/subagent-contract.md
If running through Codex, use ~/.agents/rules/.
For complex investigations, Datadog/Braintrust searches, or workflow-stage runs, read references/protocol-index.md and any local gotchas.md.
Flow
- Determine the research question from
$ARGUMENTS, the conversation, a linked ticket, or a workflow ledger. Do not ask a blank intake question when context is available.
- Read explicit files, tickets, plans, specs, and ledgers first.
- Discover relevant code with focused agents:
codebase-locator for locations only.codebase-analyzer for deep implementation reading.codebase-pattern-finder for conventions and similar implementations.requirements-tracer only for change-impact or regression-risk questions.
- Search Linear, Notion, Google Drive, and prior thought artifacts when product intent or prior decisions may live there.
- Resolve contradictions by reading primary code or source artifacts.
- Run adversarial challenge before finalizing.
- Save a research document in
~/.claude/thoughts/shared/research/NNN_topic.md.
- If a workflow ledger exists, append the research artifact path and factual assumptions.
Research Artifact
Include:
- Research question
- Executive summary
- Detailed findings with file:line evidence
- Architecture or data-flow notes
- Files and external artifacts examined
- Open questions or remaining uncertainty
Output
Return a concise summary and the research artifact path. Do not present unverified claims as findings.
