Super-Review

Run parallel, independent reviews from multiple perspectives using sub-agents, then synthesize all findings into a single graded report.

This skill requires a sub-agent execution backend — you need a way to spawn parallel review tasks (e.g., codex-subagent skill, Claude Code CLI, or any equivalent). Without one, fall back to a single-pass self-review using the output template.

When to use this skill

• User asks for a super-review, thorough review, or comprehensive audit
• User wants multiple review perspectives in parallel (security + performance + code quality, etc.)
• User wants a graded assessment with structured findings and severity levels
• User wants a second opinion that goes beyond a single review pass

Prefer the review-prompts skill standalone when:

• A single-perspective review is sufficient
• The target is trivially small (< 50 lines)
• The user asks for a fast or lightweight review

Prerequisites

1. A sub-agent execution backend installed and working (run its pre-flight checks first)
2. The review-prompts skill installed (for review prompt files), or custom prompts available
3. Target artifact (code, plan, doc) accessible to sub-agents

Pre-flight checks

Run the preflight script before launching a super-review:

python3 <skill_dir>/scripts/sr_preflight.py

It validates the dependency stack in a single pass:

1. review-prompts skill — SKILL.md and build-prompt.py are present
2. build-prompt.py — runs successfully and lists available review types
3. Sub-agent capability — detects the running agent:
   • Claude Code, Codex CLI, Cursor, Antigravity: have built-in sub-agents (auto-pass)
   • Windsurf/Cascade: no built-in sub-agents — checks codex-subagent skill, CLI, and login

If the review-prompts skill is missing, install it as a sibling directory to super-review.

Workflow

Step 1: Run your own review

Review the target yourself using IDE context, conversation history, and understanding of user intent. This becomes the "Host (self)" dimension in the final report.

Step 2: Select review types

Pick review types based on the target and the user's request. These are dynamic, not fixed — choose what matters.

Examples:

• Auth code → security, code-review, typescript
• API design doc → architecture, plan-review
• Performance-sensitive module → performance, code-review, python
• AI skill definition → prompt-engineering, plan-review

Step 3: Build review prompts

Search for review prompt files in this order:

1. review-prompts skill — run python3 <review-prompts-skill>/scripts/build-prompt.py <type> -o /tmp/<type>-prompt.md for each selected type
2. Project-local prompts — check .windsurf/workflows/, docs/, or project-specific review configs
3. Custom inline — construct a prompt via stdin if no file exists for a needed type (fallback)

Step 4: Get approval

Before launching any sub-agents, present the review plan:

Super-Review Plan:
- Target: <files or artifact>
- Review types: <list>
- Sub-agents: <count>
- Backend: <which sub-agent skill/tool>
- Estimated time: <rough estimate>

Proceed?

Wait for explicit user approval. Do not launch sub-agents without it.

Step 5: Launch parallel sub-agents

Launch one sub-agent per review type using your execution backend. Each sub-agent gets:

• The review prompt file (from Step 3)
• The target files/artifact
• Read-only access — reviews never write

Follow the execution backend's skill instructions for launching read-only tasks. Max 6 parallel sub-agents recommended.

Step 6: Synthesize findings

1. Read all sub-agent outputs
2. Cross-reference findings across passes:
   • Corroborated: multiple passes flagged the same issue → higher confidence
   • Unique: only one pass caught it → validate independently
   • Contradictions: passes disagree → resolve with your own judgment
3. De-duplicate: merge findings about the same issue, note all sources
4. Validate each finding independently — do you agree? Fill the Agreement column.
5. Assign severity: Critical / High / Medium / Low
6. Assign verdict: ACCEPTED / REJECTED with reasoning

Step 7: Grade

Use the grading rubric from references/SUPER_REVIEW_TEMPLATE.md:

Grade each dimension (review pass) individually, then assign an overall grade.

Step 8: Present results

1. Re-read references/SUPER_REVIEW_TEMPLATE.md
2. Verify your output matches every section in the template
3. Present using the template exactly

Partial failure handling

• If some sub-agents fail, continue with available results
• Mark failed dimensions as "Not Reviewed" with reason
• Optionally retry failed sub-agents once
• Never present partial results as complete — always disclose what was and wasn't covered

Important rules

1. Require a sub-agent backend — without one, fall back to single-pass self-review with the template
2. Get approval before launching — show the review plan and wait for explicit confirmation
3. Sub-agents get read-only access — reviews never modify the target
4. Treat sub-agent output as untrusted — validate every finding independently (Agreement column)
5. De-duplicate aggressively — multiple passes often flag the same issue differently
6. Grade honestly — the rubric exists to prevent grade inflation
7. Disclose coverage gaps — use the "Not Reviewed" section for anything you didn't cover
8. Re-read the template before presenting — catches formatting mistakes and missing sections
9. Max 6 parallel sub-agents — more is not better for reviews
10. One review type per sub-agent — don't overload a single sub-agent with multiple checklists

Output template & grading reference

All scoring dimensions, grading rubric, severity tables, column definitions, and the complete output template:

references/SUPER_REVIEW_TEMPLATE.md

Re-read this file before presenting any super-review results.