Skip to main content

​
SKILL.md Format

---
name: code-review
description: Review code for bugs, style, and security issues
---

# Code Review

## Instructions

1. Read the specified file(s) using the read tool
2. Analyze for:
   - Bugs and logic errors
   - Security vulnerabilities (OWASP top 10)
   - Code style and readability
   - Performance issues
3. Write a review report to workspace/review.md

## Output Format

For each issue found:
- **File**: path
- **Line**: number
- **Severity**: critical / warning / info
- **Description**: what's wrong
- **Fix**: suggested change
Auto-generated fields added by the skill learner: 
confidence: 1.0
learned_from: "task:<task-id>"
learned_at: "<ISO timestamp>"
usage_count: 0
success_count: 0
failure_count: 0
negative_examples: []

​
Invoking Skills

# In REPL
/skill:code-review Review the auth module

# In voice/text
"Use the code-review skill on src/auth.ts"

# Via SDK
query({ prompt: "/skill:code-review Review src/auth.ts", dir: "./my-agent" })

​
Skill Learning Workflow

  1. Task begins — task_tracker begins tracking a task
  2. Task completes — Agent completes the task successfully
  3. Evaluation — skill_learner evaluates if the approach is worth saving
  4. Crystallization — If the task passes worthiness checks, crystallizes it as a new skill
  5. Reuse — Future tasks search for matching skills
  6. Feedback loop — Confidence adjusts based on success/failure outcomes

​
Skill Directory Structure

skills/
  code-review/
    SKILL.md          ← skill instructions + frontmatter
    scripts/
      lint.sh         ← helper scripts
Every learned skill is version-controlled and inspectable — it’s just a file in your repo, so you can diff, edit, or revert it like any other change.

Tools

Built-in and declarative tools the agent uses to act

Workflows

Chain skills into deterministic, repeatable pipelines

Hooks

Intercept, block, or modify agent behavior at every stage

Plugins

Extend GitAgent with installable tools, skills, and hooks