Context Roles: Dev / Research / Review
Output quality in Claude Code depends heavily on the “role” you give the task. A practical pattern is to explicitly switch roles by phase, standardize the output structure, and reduce drift.
Below are three commonly used role templates (adapted from community config patterns). You can paste them into your chat as a “session preamble”.
1) Dev: ship a runnable version first
Use when:
- implementing features, fixing bugs, writing scripts, adding tests
- you want Claude to “land the change first, explain after”
Copy-paste prompt:
Context: Dev mode
Goal: Implement the feature and make it verifiable
Behavior:
- Write code first, explain after
- After changes, run minimal verification (tests/build/scripts)
- Keep changes atomic; avoid unrelated refactors2) Research: understand first, show evidence
Use when:
- exploring an unfamiliar codebase
- diagnosing a complex root cause
- you need evidence before conclusions
Copy-paste prompt:
Context: Research mode
Goal: Understand before acting
Behavior:
- Read/search before concluding
- State clarifying questions and hypotheses first
- Validate hypotheses with evidence
Output order: Findings (evidence) → Inference → Recommendations3) Review: actionable feedback by severity
Use when:
- PR review, quality review, security review
- you want a prioritized action list
Copy-paste prompt:
Context: Review mode
Goal: Quality + Security + Maintainability
Behavior:
- Read through before commenting
- Sort by severity: CRITICAL > HIGH > MEDIUM > LOW
- Provide actionable fixes (not just problems)
Output: group by file + include locationsWhen to switch roles
A stable cadence:
- unclear requirements / unfamiliar code: start in “Research”
- uncertain approach / broad impact: switch to Plan Mode (see Plan Mode)
- start implementing: switch to “Dev”
- ready to merge: switch to “Review”
To operationalize this (write into CLAUDE.md, codify as /command, or enforce via Hooks), start from Team Starter Kit.
Reference
- Related pages on this site: