| name | clean-context-review |
| description | Run an independent clean-context review for a refactor iteration, risky change, or pre-commit gate. Use when the user wants a skeptical second pass that sees only the iteration goal, diff, changed files, and test results, without implementation backstory. |
Clean Context Review
Use this skill for an explicit review gate on a risky slice, iteration slice, or pending commit.
Do not use it for:
- trivial edits
- broad codebase exploration
- normal implementation unless a review gate is wanted
Goal
Produce a skeptical review from clean context with minimal inputs only.
The reviewer should answer one core question:
behavior-preserving mode: Did this change reduce risk and improve locality without changing intended behavior?
behavior-changing mode: Did this change achieve the intended behavior with acceptable risk, scope, and test coverage?
Inputs To Prepare
- iteration name or review target
- goal
- acceptance criteria
- review mode
- changed files
- diff
- test commands
- test results
Do not pass implementation narrative, intended answer, or long history unless required for regression judgment.
Review Mode
- prefer a fresh reviewer sub-agent when the review matters
- manual review is acceptable fallback
- do not reuse reviewer context across review rounds
Prompt Template
Use the template in references/review-prompt.md.
Adapt only the slice-specific fields. Keep the rest stable.
What The Reviewer Must Focus On
- behavioral regressions
- missing or weak tests
- architectural backsliding
- coupling that remains or was introduced
- contract drift
- slices that are larger than the stated iteration goal
The reviewer should optimize for signal, not tone.
Output
Return:
- findings first, ordered by severity, with file references when possible
- open questions or assumptions
- verdict:
acceptable
acceptable with follow-ups
not acceptable
Do not choose the next workflow step. Return the verdict only.