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

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:

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: See references/phase3-subagent-prompt.md for the full template (each agent gets the same instructions with their impl number substituted).

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)
- Worktree locations: .worktrees/cookoff-impl-N/
- Test results from each implementation
- Fresh-eyes findings from Step 3
- Feasibility flags identified

The judge skill will:

1. Fill out the complete scoring worksheet for each implementation
2. Build the scorecard with integer scores (1-5, no half points)
3. Check hard gates (Fitness Δ≥2, any score=1)
4. Announce winner with rationale

Do not summarize or abbreviate the scoring. The judge skill output should be the full worksheet.

Cookoff-specific context: In cookoff, all implementations target the same design, so Fitness should be similar. A Fitness gap (Δ≥2) indicates one implementation deviated from or misunderstood the design - not a different approach choice.

Phase 5: Completion

Verification on winner:

Running final verification on winner (impl-2):
- npm test: 22/22 passing ✓
- npm run build: exit 0 ✓
- Design adherence: all requirements met ✓

Verification complete. Winner confirmed.

Winner: Use finish-branch

• Options: merge locally, create PR, keep as-is, discard

Losers: Cleanup

git worktree remove .worktrees/cookoff-impl-1
git worktree remove .worktrees/cookoff-impl-3
git branch -D <feature>/cookoff/impl-1
git branch -D <feature>/cookoff/impl-3
# Keep winner's worktree until merged

Write result.md: See references/phase5-result-template.md for the full template. Save to: docs/plans/<feature>/cookoff/result.md

Skills Orchestrated

Common Mistakes

Sharing a pre-made implementation plan

• Problem: All teams copy same code, no variation
• Fix: Each team uses writing-plans to create THEIR OWN plan from design doc

Dispatching agents in separate messages

• Problem: Serial dispatch instead of parallel
• Fix: Send ALL Task tools in a SINGLE message

Not using writing-plans + executing-plans

• Problem: Subagent implements ad-hoc
• Fix: Each subagent MUST use writing-plans then executing-plans

Skipping fresh-eyes

• Problem: Judge has no quality signal, just test counts
• Fix: Fresh-eyes on ALL survivors before comparing

Not checking for identical implementations

• Problem: Wasted compute on duplicates
• Fix: Diff implementations before fresh-eyes, flag if >95% similar

Forgetting cleanup

• Problem: Orphaned worktrees and branches
• Fix: Always cleanup losers, write result.md

Example Invocation

User: "Let's build this" (after brainstorming produced design.md)

Claude: I'm using cookoff.

How would you like to implement this design?

1. Single subagent - One agent plans and implements
2. Cookoff - 3 parallel agents, each creates own plan, pick best
   → Complexity: medium feature, touches auth
   → Recommendation: 3 implementations
3. Local - Plan and implement here

User: "2"

Claude: Spawning 3 parallel implementations...
Each will create their own implementation plan from the design.

[Phase 3: Create worktrees, dispatch ALL 3 agents in single message]
[Each agent: reads design → writes plan → implements]

Cookoff status:
- impl-1: planning → implementing 6/8 tasks
- impl-2: planning → implementing 4/7 tasks
- impl-3: planning → implementing 5/9 tasks

[All 3 complete]

[Phase 4: Diff check - implementations are different ✓]
[Fresh-eyes on all 3]

| Impl | Plan Approach | Tests | Fresh-Eyes |
|------|---------------|-------|------------|
| impl-1 | Component-first | 24/24 | 1 minor |
| impl-2 | Data-layer-first | 22/22 | 0 issues |
| impl-3 | TDD-strict | 26/26 | 2 minor |

Recommendation: impl-2 (cleanest)
User: "2"

[Phase 5: Verify winner, cleanup losers]

Winner: impl-2 ready to merge
Cleanup: 2 worktrees removed
Plans preserved: docs/plans/<feature>/cookoff/