Loop Engineering with Claude Code
and Fable 5

# Global Policies for This Project
- Always create a git worktree for experiments: git worktree add ../temp-$(date +%s)
- Run relevant tests after every code change
- Use low effort for routine tasks, high effort for planning
- Document all changes clearly
- Never execute destructive commands without confirmation
- Prefer secure, minimal changes

/init
Read the entire src/ directory and provide a high-level architecture summary.

Edit utils.py to implement caching for the get_data function.
Then execute: python -c "from utils import get_data; print(get_data())"

/goal: Refactor the authentication module for security and performance.
Success criteria (verify each iteration):
1. All unit tests pass (pytest auth/)
2. No security issues detected
3. Performance benchmark improved >=15%
4. Code style compliant and documented

Use a dedicated worktree. Maximum 5 iterations. Summarize final changes.

model: sonnet
effort: low

You are a strict code verifier. Score the output 1-10 against the provided rubric.
Only approve if score >=9. Provide specific fix suggestions otherwise.

Use the verifier subagent after every major change.
/loop every 30 seconds until goal is met or max 10 iterations reached.
Log progress to progress.md

Spawn parallel subagents:
- Researcher (haiku, low effort): Collect latest best practices
- Implementer (Fable 5): Apply changes to code
- Tester: Run full test suite

Only merge after all subagents reach consensus.

# Automatic Code Review Skill
Triggers: File edits in src/ or PR creation

Instructions:
- Review for tests, edge cases, security, performance, readability
- Suggest improvements in JSON format: {"issues": [...], "suggestions": [...]}
- Auto-apply safe fixes when confidence >80%

const { execSync } = require('child_process');
try {
  execSync(`eslint --fix ${process.env.FILE}`);
  console.log(`Linting applied to ${process.env.FILE}`);
} catch (e) {
  console.log('Linting warnings present');
}

Connect via MCP to GitHub.
Create a draft PR for the current branch.
Include:
- Clear title and description
- Test results summary
- Security checklist

Refine this voice transcript into a complete executable /goal loop with detailed rubric and verifier instructions.

[PASTE TRANSCRIPT HERE]

/goal: Review, fix, and create draft PR for branch 'feature/new-feature'.
Loop steps:
1. Checkout in isolated worktree
2. Run full tests + linter + security scan
3. Analyze changes and generate improvements
4. Create/update draft PR with detailed summary, test report, and checklist
5. Iterate fixes up to 4 times

Use Fable 5 for orchestration.

/goal: Create a comprehensive Q3 market analysis report on AI coding assistants.
Requirements:
- Cover major players with citations
- Include SWOT analysis and recommendations
- Professional Markdown format with tables
- Verify all facts across multiple sources

Run as a daily Routine.

Create a detailed blog post on "Loop Engineering Best Practices".
Process: Outline → Draft sections with voice input → Self-critique → Revise → Format for publication.

Process leads from leads.csv:
- Personalize outreach emails using template
- Send via approved MCP integration
- Track responses in loop
- Generate daily summary report

main: fable-5
workers:
  - name: scout
    model: haiku
    effort: low
  - name: executor
    model: sonnet

# Loop Progress
Current Goal: ...
Attempts: ...
Learnings: ...
Next Actions: ...