mohammadhprp/teammate/review-change
Claude Code and agent skill by mohammadhprp/teammate
- community
Checking that the SKILL.md file is still in the repository.
Review a diff against its acceptance criteria with fresh evidence: read the full changed files, run the project's own checks, record one actionable finding per observation with severity and category, and give a pass/fail/inconclusive verdict that names the checks run. Use when a completed change must be judged against agreed criteria — even if the request only says 'check this' or 'does this look right'.
Install
npx vskill@latest install mohammadhprp/teammate/review-changeSKILL.md
The same SKILL.md was published earlier as mohammadhprp/teammate/review-change.
Review change
Review the work, not the summary. A report is a claim; the diff and the project's own checks are the evidence.
This is the shared, read-only method. It changes nothing: no fixes, no commits, no pushes, no comments on a remote. verify-evidence defines what counts as evidence.
When to use
- A completed change — a local diff, a branch, or a patch — must be checked
against acceptance criteria.
- The primary must verify a change before approval.
When not to use
- The task is still running: the primary waits for it to return
(monitor-agents).
- The assignment includes fixing the change: that is implementation, not review.
- You are deciding whether to review, delegating a reviewer, or driving rework:
that coordination is the primary's (review-work), not this method.
Inputs
- The goal and acceptance criteria for the change.
- The change itself: a local diff, a branch, or a patch.
- The project's check commands (from
load-project-context).
Fresh evidence
Collect evidence now. Never review from memory or reuse a diff snapshot from an earlier iteration; the working tree may have moved.
- Establish intent: the goal, every acceptance criterion, and the constraints
from the request or brief. If criteria are missing, derive them from the goal and say what you assumed.
- Read the diff —
git diffandgit diff --staged; for another agent's
working copy use git -C <root> diff. As the primary, python3 scripts/tm.py diff --cwd <root> is also available; it lists untracked files, so read a new file's contents directly.
- Read each changed file in full, not only the hunks; surrounding context
decides whether a change is correct.
- Run the project's own checks from its
AGENTS.md/CONTEXT.md: tests,
linters, type checks. Record the exact command and result.
- Judge against the project's
AGENTS.mdconventions and acceptance bar, not
just the presence of code. For a user-facing result, inspect the rendered or visible output — a screenshot or the built page — not only its source text.
- Walk each acceptance criterion and mark it demonstrated, violated, or
unverified.
Findings
Record one observation per finding with a severity, category, title, detail, file, line, suggestion, and status. The schema, categories, severity levels, and verdict rules live in references/findings.md.
- Include only actionable findings. A concern about untouched code belongs in
the verdict notes, not as a finding on the change.
- Treat a pasted command transcript as a claim, not evidence: re-run the
decisive check yourself. A transcript with no run behind it, or one that does not match a fresh run, is a blocker finding for fabricated evidence.
- Do not pad the list with speculation or style preferences; noise buries the
real defects.
- If generated or binary noise pollutes the diff, treat it as a project hygiene
finding rather than blaming the author by default.
Verdict
Name the checks you actually ran; the verdict rests on them.
- pass — every acceptance criterion holds and no open
blocker/major
remains. A pass with zero checks is a process failure, not a pass.
- fail — at least one open
blocker/major. - inconclusive — correctness cannot be determined, usually because evidence
is missing or the change cannot be run.
When correctness cannot be determined, record inconclusive rather than forcing a pass or a fail; a guessed verdict is worse than an honest gap. Record it on the build task and escalate to the developer:
pass and fail stay derived from the findings; --verdict auto clears an inconclusive back to that derived value.
Output
A findings list and a verdict (pass / fail / inconclusive) that names the criteria and the checks run. Return findings only; do not fix, commit, or approve.
Failure and escalation
- The project's checks cannot run, or evidence is missing — report
inconclusive.
inconclusive, a disputed finding, or a criterion that cannot be interpreted
— hand it back to whoever requested the review instead of resolving it silently. As the primary that is review-work; as a reviewer worker it is the brief's requester.
- Never approve. Approval is the developer's decision, not the reviewer's.
Security
- Security scanPassed, 90 of 100Pattern checks on the skill's files.
- Model reviewPassed
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryCheckingsrc/skills/review-change/SKILL.md
Popularity
- 7-day trend score
- 0
Not refreshed yet.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/mohammadhprp/teammate/review-change)