cookoff

Cookoff

Same design, multiple cooks compete. Each implementation team creates their own plan from the shared design, then implements it. Natural variation emerges from independent planning decisions.

Part of Test Kitchen Development:

• omakase-off - Chef's choice exploration (different approaches/designs)
• cookoff - Same design, multiple cooks compete (each creates own plan + implements)

Key insight: Don't share a pre-made implementation plan. Each agent generates their own plan from the design doc, ensuring genuine variation.

Directory Structure

docs/plans/<feature>/
  design.md                    # Input: from brainstorming
  cookoff/
    impl-1/
      plan.md                  # Agent 1's implementation plan
    impl-2/
      plan.md                  # Agent 2's implementation plan
    impl-3/
      plan.md                  # Agent 3's implementation plan
    result.md                  # Cookoff results and winner

Skill Dependencies

| Reference | Primary (if installed) | Fallback | |-----------|------------------------|----------| | writing-plans | superpowers:writing-plans | Each agent writes their own implementation plan | | executing-plans | superpowers:executing-plans | Execute plan tasks sequentially with verification | | parallel-agents | superpowers:dispatching-parallel-agents | Dispatch multiple Task tools in single message | | git-worktrees | superpowers:using-git-worktrees | git worktree add .worktrees/<name> -b <branch> | | tdd | superpowers:test-driven-development | RED-GREEN-REFACTOR cycle | | verification | superpowers:verification-before-completion | Run command, read output, THEN claim status | | fresh-eyes | fresh-eyes-review:skills (2389) | 2-5 min review for security, logic, edge cases | | judge | test-kitchen:judge | Scoring framework with checklists (MUST invoke at Phase 4) | | code-review | superpowers:requesting-code-review | Dispatch code-reviewer subagent | | scenario-testing | scenario-testing:skills (2389) | .scratch/ E2E scripts, real dependencies | | finish-branch | superpowers:finishing-a-development-branch | Verify tests, present options, cleanup |

When to Use

Trigger when user wants to implement a design:

• "Execute this plan" / "Implement the plan" / "Let's build this"
• After brainstorming completes and design doc exists
• Can also invoke explicitly: "cookoff this"

Important: Cookoff works from a design doc, not a detailed implementation plan. Each agent creates their own implementation plan.

Detecting the Design-to-Implementation Transition

Cookoff triggers at a SITUATION, not a specific skill's output.

The situation: Design is complete, implementation is about to start.

Signals that design phase just completed:

• Design doc was written/committed
• User approved a design ("looks good", "yes", "let's do it")
• Discussion shifted from "what to build" to "how to build it"
• Any skill/flow is about to start implementation

When you detect this transition, ALWAYS offer cookoff:

Before we start implementation, how would you like to proceed?

1. Cookoff (recommended) - N parallel agents, each creates own plan, pick best
   → Complexity: [assess from design]
   → Best for: medium-high complexity features
2. Single implementation - One agent/session implements
3. Direct coding - Start coding without detailed plan

This applies regardless of:

• Which brainstorming skill was used (superpowers, other, or none)
• Whether a formal design doc exists (could be informal agreement)
• What implementation options another skill might present

The key insight: We're not injecting into another skill's menu. We're recognizing a SITUATION (design→implementation) and ensuring cookoff is offered at that moment.

Phase 1: Implementation Options

Present choices when user wants to implement:

How would you like to implement this design?

1. Single subagent - One agent plans and implements
2. Cookoff - N parallel agents, each creates own plan, pick best
   → Complexity: [assess from design]
   → Recommendation: N implementations
3. Local - Plan and implement here in this session

Which approach?

Routing:

• Option 1: Single agent uses writing-plans then executing-plans, cookoff exits
• Option 2: Continue to Phase 2
• Option 3: User implements manually, cookoff exits

Phase 2: Complexity Assessment

Read design doc and assess:

• Feature scope (components, integrations, data models)
• Risk areas (auth, payments, migrations, concurrency)
• Estimated implementation size

Map to implementation count:

| Complexity | Scope | Risk signals | Implementations | |------------|-------|--------------| --------------- | | Low | Small feature | None | 2 | | Medium | Medium feature | Some | 3 | | High | Large feature | Several | 4 | | Very high | Major system | Critical areas | 5 |

Setup directories:

mkdir -p docs/plans/<feature>/cookoff/impl-{1,2,3}

Announce:

Complexity assessment: medium feature, touches auth
Spawning 3 parallel implementations
Each will create their own implementation plan from the design.

Phase 3: Parallel Execution

Setup worktrees:

.worktrees/cookoff-impl-1/
.worktrees/cookoff-impl-2/
.worktrees/cookoff-impl-3/

Branches:
<feature>/cookoff/impl-1
<feature>/cookoff/impl-2
<feature>/cookoff/impl-3

CRITICAL: Dispatch ALL agents in a SINGLE message

Use parallel-agents pattern. Send ONE message with multiple Task tool calls:

<single message>
  Task(impl-1, run_in_background: true)
  Task(impl-2, run_in_background: true)
  Task(impl-3, run_in_background: true)
</single message>

Do NOT send separate messages for each agent.

Subagent prompt (each gets same instructions with their impl number):

You are implementation team N of M in a cookoff competition.
Other teams are implementing the same design in parallel.
Each team creates their own implementation plan - your approach may differ from others.

**Your working directory:** /path/to/.worktrees/cookoff-impl-N
**Design doc:** docs/plans/<feature>/design.md
**Your plan location:** docs/plans/<feature>/cookoff/impl-N/plan.md

**Your workflow:**
1. Read the design doc thoroughly
2. Use writing-plans skill to create YOUR implementation plan
   - Save to: docs/plans/<feature>/cookoff/impl-N/plan.md
   - Make your own architectural decisions
   - Don't try to guess what other teams will do
3. Use executing-plans skill to implement your plan
4. Follow TDD for each task
5. Use verification before claiming done

**Report when done:**
- Plan created: yes/no
- All tasks completed: yes/no
- Test results (npm test output)
- Files changed count
- Any issues encountered

Your goal: best possible implementation. Good luck!

Monitor progress:

Cookoff status (design: auth-system):
- impl-1: planning... → implementing 5/8 tasks
- impl-2: planning... → implementing 3/8 tasks
- impl-3: planning... → implementing 6/8 tasks

Phase 4: Judging

Step 1: Gate check

• All tests pass
• Design adherence - implemented what the design specified

Step 2: Check for identical implementations

Before fresh-eyes, diff the implementations:

diff -r .worktrees/cookoff-impl-1/src .worktrees/cookoff-impl-2/src

If implementations are >95% identical, note this - the planning step didn't create enough variation. Still proceed but flag in results.

Step 3: Fresh-eyes on survivors

Starting fresh-eyes review of impl-1 (N files)...
Checking: security, logic errors, edge cases
Fresh-eyes complete: 1 minor issue

Step 4: Invoke Judge Skill

CRITICAL: Invoke test-kitchen:judge now.

The judge skill contains the full scoring framework with checklists. Invoking it fresh ensures the scoring format is followed exactly.

Invoke: test-kitchen:judge

Context to provide:
- Implementations to judge: impl-1, impl-2, impl-3 (or however many)
- Workt