meld-adversarial-review

Category: Development & Coding | Uploader: solomonjames | Downloads: 0 | Version: v1.0（Latest）

Use when implementation is complete and you need a code review with information asymmetry — reviewer sees ONLY the diff, not the spec or conversation history

Changelog: Source: GitHub https://github.com/solomonjames/meld

Directory Structure

Current level: skills/meld-adversarial-review/

📁 evals/
- 📁 eval-files/
  - 📄 diff-1-sql-injection.diff 2.7 KB
  - 📄 diff-2-react-component.diff 3.8 KB
  - 📄 diff-3-clean-utility.diff 3.5 KB
  - 📄 diff-4-multifile-auth.diff 5.7 KB
- 📁 evals/
  - 📄 evals.json 4.9 KB
- 📁 iteration-1/
  - 📁 eval-1-sql-injection-vuln/
    
    📁 with_skill/
    
    📁 outputs/
    
    📄 review_output.md 5.2 KB
    
    📄 reviewer_raw_findings.md 3.4 KB
    
    📄 severity_validity_assessment.md 2.1 KB
    
    📄 grading.json 1.6 KB
    
    📄 timing.json 88 B
    
    📁 without_skill/
    
    📁 outputs/
    
    📄 review_output.md 5.4 KB
    
    📄 grading.json 1.6 KB
    
    📄 timing.json 86 B
    
    📄 eval_metadata.json 709 B
  - 📁 eval-2-react-xss/
    
    📁 with_skill/
    
    📁 outputs/
    
    📄 review_output.md 5.5 KB
    
    📄 grading.json 1.6 KB
    
    📄 timing.json 86 B
    
    📁 without_skill/
    
    📁 outputs/
    
    📄 review_output.md 4.6 KB
    
    📄 grading.json 1.6 KB
    
    📄 timing.json 85 B
    
    📄 eval_metadata.json 703 B
  - 📁 eval-3-clean-cache/
    
    📁 with_skill/
    
    📁 outputs/
    
    📄 review_output.md 4.7 KB
    
    📄 grading.json 1.4 KB
    
    📄 timing.json 86 B
    
    📁 without_skill/
    
    📁 outputs/
    
    📄 review_output.md 5.5 KB
    
    📄 grading.json 1.3 KB
    
    📄 timing.json 85 B
    
    📄 eval_metadata.json 623 B
  - 📄 benchmark.json 8.3 KB
  - 📄 feedback.json 982 B
  - 📄 review.html 97.8 KB
- 📁 iteration-2/
  - 📁 eval-1-sql-injection-vuln/
    
    📁 with_skill/
    
    📁 outputs/
    
    📄 review_output.md 6.4 KB
    
    📄 grading.json 1.3 KB
    
    📄 timing.json 86 B
    
    📁 without_skill/
    
    📁 outputs/
    
    📄 review_output.md 5.2 KB
    
    📄 grading.json 1.2 KB
    
    📄 timing.json 86 B
    
    📄 eval_metadata.json 952 B
  - 📁 eval-2-react-xss/
    
    📁 with_skill/
    
    📁 outputs/
    
    📄 review_output.md 4.5 KB
    
    📄 grading.json 1.2 KB
    
    📄 timing.json 86 B
    
    📁 without_skill/
    
    📁 outputs/
    
    📄 review_output.md 4.9 KB
    
    📄 grading.json 1.2 KB
    
    📄 timing.json 85 B
    
    📄 eval_metadata.json 946 B
  - 📁 eval-3-clean-cache/
    
    📁 with_skill/
    
    📁 outputs/
    
    📄 review_output.md 3.5 KB
    
    📄 grading.json 1.1 KB
    
    📄 timing.json 86 B
    
    📁 without_skill/
    
    📁 outputs/
    
    📄 review_output.md 5.4 KB
    
    📄 grading.json 1.1 KB
    
    📄 timing.json 85 B
    
    📄 eval_metadata.json 866 B
  - 📁 eval-4-multifile-auth/
    
    📁 with_skill/
    
    📁 outputs/
    
    📄 review_output.md 5.4 KB
    
    📄 grading.json 1.3 KB
    
    📄 timing.json 86 B
    
    📁 without_skill/
    
    📁 outputs/
    
    📄 review_output.md 8.5 KB
    
    📄 grading.json 1.1 KB
    
    📄 timing.json 86 B
    
    📄 eval_metadata.json 986 B
  - 📄 benchmark.json 10.3 KB
  - 📄 review.html 147.7 KB
- 📁 skill-snapshot/
  - 📄 reviewer-prompt.md 1.6 KB
  - 📄 SKILL.md 2.5 KB
📄 reviewer-prompt.md 1.6 KB
📄 SKILL.md 4.0 KB

SKILL.md

---
name: meld-adversarial-review
description: Use when implementation is complete and you need a code review with information asymmetry — reviewer sees ONLY the diff, not the spec or conversation history
---

# Adversarial Review

This skill performs a code review with **information asymmetry**: the reviewer sees ONLY the diff, not the spec, conversation history, or implementation intent. This simulates a fresh reviewer who judges purely on code quality.

## Prerequisites

- A baseline commit (captured before implementation started)
- Implementation is complete and self-check has passed

## Procedure

### 1. Construct the Diff

**Git available:**
```bash
git diff {baseline_commit}
```
- Include both tracked and untracked new files
- For new files: `git diff --no-index /dev/null {file}` or include full content

**No git fallback:**
- List all files modified/created during the workflow
- Include full content of changed files for review

**Large diffs (>500 lines):** If the diff is very large, note this when passing it to the reviewer. The reviewer should prioritize security and correctness findings over style issues.

### 2. Invoke the Reviewer

**REQUIRED:** Use the Agent tool to spawn a subagent for true information asymmetry. The subagent must receive ONLY the diff and the reviewer prompt — no spec, no conversation history, no implementation context.

Read the reviewer prompt from `reviewer-prompt.md` in this skill directory, then construct the subagent call like this:

```
Agent tool call:
  prompt: "<contents of reviewer-prompt.md>\n\n<diff>\n<paste the full diff content here>\n</diff>"
  subagent_type: "general-purpose"
  description: "Adversarial code review"
```

The key constraint: the subagent's prompt must contain ONLY the reviewer instructions and the diff. Nothing else. This is what creates the information asymmetry — the reviewer cannot be biased by knowing what the code is supposed to do.

**Fallback (only if Agent tool is genuinely unavailable):** Perform the review inline. Before reviewing, explicitly discard all prior context. Start fresh: read only the diff, apply only the reviewer prompt criteria. State clearly in your output that this is an inline review without true information asymmetry.

### 3. Process Findings

For each finding from the reviewer, evaluate two dimensions:

**Severity** (how bad is it?):
- **Critical:** Security vulnerability, data loss risk, crash
- **High:** Incorrect behavior, missing error handling, test gap
- **Medium:** Code smell, performance concern, maintainability issue
- **Low:** Style inconsistency, minor improvement opportunity

**Validity** (is the reviewer right, given that they lack context?):
- **Real:** Genuine issue that should be fixed
- **Noise:** False positive — the reviewer lacks context that makes this a non-issue
- **Undecided:** Could go either way, needs human judgment

### 4. Present Findings

Format findings with a numbered ID, severity, validity, and location:

```
F1 [Critical/Real] — Description of finding
   Explanation of why this matters.
   Location: file:line

F2 [High/Undecided] — Description of finding
   Explanation of why this matters.
   Location: file:line
```

Order by severity (Critical first, Low last). Within the same severity, list Real before Undecided before Noise.

After the findings list, include a **summary table**:

| Severity | Count | Real | Undecided | Noise |
|----------|-------|------|-----------|-------|
| Critical | N | N | N | N |
| High | N | N | N | N |
| Medium | N | N | N | N |
| Low | N | N | N | N |
| **Total** | **N** | **N** | **N** | **N** |

### 5. Zero Findings Check

If the review produces zero findings:
- Flag this as suspicious
- Consider: was the diff too small? Was the review thorough?
- Note the anomaly but proceed

## Resolution

Present resolution options to the user:
- **[W] Walk through** — Iterate each finding, ask Fix/Skip/Discuss for each
- **[F] Fix automatically** — Apply fixes for "Real" findings, skip Noise/Undecided
- **[S] Skip** — Acknowledge findings and proceed without fixes

Comments 0

Latest | Hot

Please login before commenting.

No comments yet. Be the first one!

meld-adversarial-review

Directory Structure

SKILL.md

Report

Notice