# Compare Parallel Implementations

Teams often lose quality signals when one agent implements a design alone. This skill runs parallel implementation trials from the same design and guides comparison, review, and cleanup.

## Install

```bash
npx skillstore add 2389-research/cookoff
```

## Metadata

- - Status: approved
- - Slug: 2389-research-cookoff
- - Version: 1.0.0
- - Author: 2389-research
- - GitHub username: 2389-research
- - License: MIT
- - Repository: https://github.com/2389-research/claude-plugins/tree/main/test-kitchen/skills/cookoff
- - Ref: main
- - Supported tools: Claude, Codex, Claude Code
- - Risk level: medium
- - Risk factors: filesystem, external\_commands
- - Quality score: 73
- - Public page: https://skillstore.pages.dev/skills/2389-research-cookoff
- - Manifest: https://skillstore.pages.dev/api/skills/2389-research-cookoff/manifest

## Capabilities

- Detects when a design is ready for implementation.
- Offers single-agent, parallel-agent, and local implementation paths.
- Maps feature complexity to a suggested number of implementations.
- Defines a worktree-based structure for independent implementation attempts.
- Guides review, judging, winner selection, and cleanup.
- Preserves plans and result summaries under docs/plans.

## Use Cases

- Compare architecture choices: Run several implementations from one design to compare maintainability, test coverage, and design fit.
- Reduce implementation risk: Use independent plans and fresh reviews before choosing code for a risky feature.
- Structure agent coding work: Give Claude, Codex, or Claude Code a repeatable process for parallel coding trials.

## Prompt Templates

### Start from an approved design

```
Use cookoff for this approved design. Show my implementation options before creating worktrees or dispatching agents.
```

### Run a small comparison

```
Use cookoff with two implementations. Keep both teams focused on the same design and report the differences after tests pass.
```

### Coordinate a high-risk feature

```
Use cookoff for this feature that touches authentication. Spawn independent plans, require TDD, run fresh review, and judge the survivors.
```

### Select and clean up a winner

```
Finish the cookoff by verifying the winner, writing the result summary, and asking before removing losing worktrees or branches.
```

## Limitations

- Requires a clear design or implementation goal before use.
- Depends on compatible agent and git worktree workflows.
- Does not implement code by itself without an execution agent.
- Cleanup commands must be reviewed before execution.

## Best Practices

- Start with a design document or a clear approved scope.
- Let each implementation team create its own plan from the same design.
- Review cleanup commands before deleting worktrees or branches.

## Anti Patterns

- Do not give every team the same detailed implementation plan.
- Do not dispatch parallel teams one at a time.
- Do not select a winner using test counts alone.

## Security Audit

- - Safe to publish: true
- - Audited at: 2026-06-27T16:24:10.59\+00:00
- - Summary: The static analyzer reported many command, filesystem, weak cryptography, and reconnaissance patterns. Review found no executable malware, cryptographic code, prompt injection, or data exfiltration; most high-risk alerts are false positives from prose and examples. The remaining risk is legitimate but material because the skill instructs agents to create worktrees, run shell commands, write files, and remove git branches.

## Stats

- - Views: 275
- - Downloads: 11
- - Favorites: 0
- - Popularity score: 0