Review Skill

Intent

Provide prioritized findings on correctness, safety, and maintainability with evidence.

Trigger

Use this skill for:

• •code review requests
• •pre-merge risk checks
• •regression or quality audits

Review Workflow

Step 1: Build review context

Collect:

• •changed files and intent
• •related tests and verification signals
• •impacted interfaces/contracts

If no diff is available, review requested target files with explicit assumptions.

Step 2: Inspect by dimension

Evaluate every target across these dimensions:

1. •Correctness and edge cases
2. •Regression risk and compatibility
3. •Security and data handling
4. •Performance/resource implications
5. •Maintainability and clarity

Step 3: Severity classification

Use consistent severity labels:

• •P0: release-blocking, data loss, security-critical
• •P1: high-impact functional bug or likely regression
• •P2: medium issue affecting reliability or maintainability
• •P3: low-impact improvement opportunity

Each finding must include:

• •why it is a problem
• •where it appears (file/symbol)
• •expected impact
• •concrete fix direction

Step 4: Emit findings first

Output order:

1. •findings (highest severity first)
2. •open questions/assumptions
3. •short summary

Template:

FINDING
- severity: <P0|P1|P2|P3>
- location: <path:symbol>
- issue: "<what is wrong>"
- impact: "<user/system impact>"
- recommendation: "<specific fix>"

Step 5: Report testing gaps

List missing or weak verification coverage:

• •absent unit tests for new branches
• •missing integration checks for shared boundaries
• •no validation for failure/recovery paths

Review Heuristics

• •Favor root-cause findings over style nits.
• •Prioritize behavior-changing issues over formatting.
• •Treat mutable shared state and concurrency as high-risk.
• •Treat contract/schema drift as high-risk even if tests pass locally.

Stop Conditions

• •Cannot identify intent or scope of reviewed changes.
• •Required files are missing from repository snapshot.
• •Evidence is insufficient to make reliable severity call.

When blocked, state exactly which artifact is missing.

Anti-Patterns (never)

• •Giving only generic praise without findings.
• •Reporting style-only comments as top findings.
• •Claiming "no issues" without discussing residual risk.
• •Mixing uncertain speculation with confirmed bugs.

Example

Input:

"Review verification gate changes for regressions and missing tests."

Expected output:

1. •Findings list with P0-P3 severity.
2. •Assumptions and open questions.
3. •testing_gaps section for missing coverage.