Team review
Reply in the owner's language (profile; default Egyptian Arabic).
Parallel multi-angle review, then adversarial verification of every finding before showing it.
$ARGUMENTS = what to review. Empty -> latest changes.
1. Define scope first
git diff --stat HEAD # uncommitted changes
git diff --stat main...HEAD # whole branch
- Diff empty and no
$ARGUMENTS? Stop and ask viaAskUserQuestionwhat to review. Never review the whole repo. - Under ~3 files and a simple change? Review it yourself directly. Don't launch a team: coordination cost exceeds the benefit. Tell the owner in one line that you chose direct review and why.
2. Pick angles by what actually changed
Don't always launch all four. Each angle has an activation condition:
| Angle | agent | Activate when |
|---|---|---|
| Correctness and safety | code-reviewer |
always |
| Security | security-reviewer |
auth · secrets · user input · file paths · external requests · payments · personal data |
| Tests | qa-tester |
new or modified logic with branches |
| UI | ui-ux (review mode) |
.tsx/.jsx/.vue/CSS files or anything the user sees |
| Infrastructure | devops-sre |
Dockerfile · CI · k8s · deploy config |
3. Launch in parallel
All calls in one message so they truly run in parallel.
Agent(subagent_type: "<name>", run_in_background: true,
prompt: "Review this diff from your angle. Scope: <files>.
Report everything you find, including low-confidence and low-severity,
with confidence and severity for each finding. Do not filter yourself.")
Kit agents take their model from their own frontmatter: do not pass model to them. Built-in agents (Explore) get model: "sonnet" explicitly.
Use SendMessage to ask a teammate for clarification; it resumes with full context.
4. ⚠️ Verify adversarially: the step that makes the difference
Never show reviewers' findings as-is. A reviewer is steered toward coverage, not precision, so it will produce false positives by design. Filtering is your job.
For each significant finding, launch a verifier whose task is refutation:
Agent(subagent_type: "Explore", model: "sonnet",
prompt: "Try to refute this claim: '<finding>'.
Read the actual code. Is the stated path really reachable?
Is there a guard the reviewer missed?
Default: REFUTED if you cannot prove the path.")
Show only what survived. State how many findings fell: that number tells the owner the review's own quality.
5. Report
Owner-facing template (headings stay in Arabic):
## 🔍 مراجعة الفريق — <النطاق>
**الزوايا:** <المُفعَّلة> · **النتائج:** <مؤكَّدة> مؤكَّدة من <إجمالي> (سقطت <n> في التحقق)
### 🔴 يمنع الشحن
- `file.ts:42` — <ما ينكسر> · **يحدث حين:** <مدخل أو حالة محددة> · **الإصلاح:** <جملة>
### 🟡 يجب إصلاحه
### 🟢 يستحق النظر
### ✅ ما فُحص ووجد سليماً
<حتى لا يُعاد فحصه>
(Template legend: scope · angles activated · confirmed of total (n dropped in verification) · blocks shipping: what breaks / happens when: specific input or state / fix: one sentence · should fix · worth considering · checked and found sound, so it isn't rechecked.)
Describe the fix, don't write the corrected code, unless explicitly permitted (CLAUDE.md §2).
No findings? Say so in one line. Don't manufacture findings to look useful.
When not to use this skill
- One-line change or typo fix
- Code not yet written: review after implementation
- The owner is asking a question about the code, not requesting a review