Adversarial Reviewer

SKILL-019: Adversarial Reviewer

Overview

Forces a context switch from "Builder" to "Attacker". Ideally used before finalizing any critical component (e.g. auth, payments, file I/O). It prepares a structured prompt packet that the agent then uses to critique its own work.

Trigger Phrases

• red team this
• adversarial review
• find bugs
• critique code

Inputs

Outputs

1. Structured Prompt (Markdown)

A prompt template pre-filled with the code and specific attack vectors. The Agent must then "simulate" the adversary by responding to this prompt.

# ⚔️ ADVERSARIAL REVIEW REQUEST
**TARGET:** src/Auth.cs
**MODE:** SECURITY

## INSTRUCTIONS
You are now the ADVERSARY. Break this code.
Look for:
1. Race Conditions
2. Replay Attacks
...

Preconditions

1. Target file exists.
2. PowerShell 5.1+ or Core 7+.

Safety/QA Checks

1. Read-Only: Does not modify the code; only reads it.
2. Size Limit: If file is massive (>10k lines), warns or truncates to prevent token overflow.

Stop Conditions

Implementation

See scripts/prepare_review.ps1.

Integration with Other Skills

1. Call prepare_review.ps1.
2. Agent Step: Read the output text.
3. Agent Step: "Thinking" block -> Process the critique.
4. Agent Step: Generate list of fixes.