agentleFS
2389-research2389-research/speed-run/skills/showdown/SKILL.md

showdown

Skill1 starsChanged 4 months agoMITView on GitHub

Runs 2-5 parallel runners against a shared design doc in isolated git worktrees, each independently planning and generating code via Cerebras hosted LLM, then invokes speed-run:judge to score all implementations and pick the winner (~60% token savings vs Claude direct). Use when you have a finalized design and want the best implementation rather than architectural exploration.

1 other project uses a version of this.

Adopted by

Most recently taken up by

  • Other versions1 project uses a version of this.

    Word for word · 0

    None found.

    Edited · 0

    None found.

    Wrote their own version · 1

    1. rubicon/ai-skills★ 1same name

    How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.

  • In this repository5 other files it gives its agents.

The file

We split this skill into the rules it gives an agent

5 rules. About 0 are common, 4 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈4

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Key insight: Don't share a pre-made implementation plan.

    line 12
  2. Very specific:

    Use parallel-agents pattern.

    line 85
  3. Distinctive:

    Before fresh-eyes, diff the implementations:

    line 113
  4. Distinctive:

    If implementations are >95% identical, note this - the planning step didn't create enough variation.

    line 118
  5. Distinctive:

    Check hard gates (Fitness Δ≥2, any score=1)

    line 148

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Showdown

Same design, multiple runners compete. Each runner creates their own implementation plan from the shared design, then generates code via hosted LLM. Natural variation emerges from independent planning decisions.

Announce: "I'm using speed-run:showdown for parallel competition via hosted LLM."

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

Directory Structure

docs/plans/<feature>/
  design.md                    # Input: from brainstorming
  speed-run/
    showdown/
      runner-1/
        plan.md                # Runner 1's implementation plan
      runner-2/
        plan.md                # Runner 2's implementation plan
      runner-3/
        plan.md                # Runner 3's implementation plan
      result.md                # Showdown results and winner

Skill Dependencies

ReferencePrimary (if installed)Fallback
writing-planssuperpowers:writing-plansEach runner writes their own plan
executing-planssuperpowers:executing-plansExecute plan tasks sequentially
parallel-agentssuperpowers:dispatching-parallel-agentsDispatch multiple Task tools in single message
git-worktreessuperpowers:using-git-worktreesgit worktree add .worktrees/<name> -b <branch>
tddsuperpowers:test-driven-developmentRED-GREEN-REFACTOR cycle
verificationsuperpowers:verification-before-completionRun command, read output, THEN claim status
fresh-eyesfresh-eyes-review:fresh-eyes-review (2389)2-5 min review for security, logic, edge cases
judgespeed-run:judgeScoring framework with checklists (MUST invoke at Phase 4)
scenario-testingscenario-testing:scenario-testing (2389).scratch/ E2E scripts, real dependencies
finish-branchsuperpowers:finishing-a-development-branchVerify tests, present options, cleanup

Phase 1: Complexity Assessment

Read design doc and assess:

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

Map to runner count:

ComplexityScopeRisk signalsRunners
LowSmall featureNone2
MediumMedium featureSome3
HighLarge featureSeveral4
Very highMajor systemCritical areas5

Announce:

Complexity assessment: medium feature, touches auth
Spawning 3 parallel runners
Each will create their own implementation plan from the design.
All runners will use hosted LLM (Cerebras) for code generation.

Phase 2: Parallel Execution

Setup worktrees:

.worktrees/speed-run-runner-1/
.worktrees/speed-run-runner-2/
.worktrees/speed-run-runner-3/

Branches:
<feature>/speed-run/runner-1
<feature>/speed-run/runner-2
<feature>/speed-run/runner-3

CRITICAL: Dispatch ALL runners in a SINGLE message

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

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

Runner prompt: see references/runner-agent-prompt.md

Monitor progress:

Showdown status (design: auth-system):
- runner-1: planning... -> generating via Cerebras -> fixing 2/3 -> tests passing
- runner-2: planning... -> generating via Cerebras -> tests passing
- runner-3: planning... -> generating via Cerebras -> fixing 1/2 -> tests passing

Phase 3: 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/speed-run-runner-1/src .worktrees/speed-run-runner-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 runner-1 (N files)...
Checking: security, logic errors, edge cases
Fresh-eyes complete: 1 minor issue

Step 4: Invoke Judge Skill

CRITICAL: Invoke speed-run:judge now.

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

Invoke: speed-run:judge

Context to provide:
- Implementations to judge: runner-1, runner-2, runner-3
- Worktree locations: .worktrees/speed-run-runner-N/
- Test results from each runner
- Fresh-eyes findings from Step 3
- Speed-run metrics: hosted LLM calls, fix cycles, generation time per runner

The judge skill will:

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

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

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

Phase 4: Completion

Verification on winner:

Running final verification on winner (runner-2):
- Tests: 22/22 passing
- Build: exit 0
- Design adherence: all requirements met

Verification complete. Winner confirmed.

Winner: Use finish-branch

Losers: Cleanup

git worktree remove .worktrees/speed-run-runner-1
git worktree remove .worktrees/speed-run-runner-3
git branch -D <feature>/speed-run/runner-1
git branch -D <feature>/speed-run/runner-3

Write result.md:

# Showdown Results: <feature>

## Design
docs/plans/<feature>/design.md

## Runners
| Runner | Plan Approach | Tests | Fresh-Eyes | Lines | LLM Calls | Fix Cycles | Result |
|--------|---------------|-------|------------|-------|-----------|------------|--------|
| runner-1 | Component-first | 24/24 | 1 minor | 680 | 4 | 2 | eliminated |
| runner-2 | Data-layer-first | 22/22 | 0 issues | 720 | 3 | 1 | WINNER |
| runner-3 | TDD-strict | 26/26 | 2 minor | 590 | 5 | 3 | eliminated |

## Plans Generated
- runner-1: docs/plans/<feature>/speed-run/showdown/runner-1/plan.md
- runner-2: docs/plans/<feature>/speed-run/showdown/runner-2/plan.md
- runner-3: docs/plans/<feature>/speed-run/showdown/runner-3/plan.md

## Winner Selection
Reason: Clean fresh-eyes review, solid data-layer-first architecture, fewest fix cycles

## Token Savings
Estimated savings vs Claude direct: ~60% on code generation

Save to: docs/plans/<feature>/speed-run/showdown/result.md

Common Mistakes

Sharing a pre-made implementation plan

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

Dispatching runners in separate messages

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

Using Claude direct for all code generation

  • Problem: Defeats the purpose of speed-run (token savings)
  • Fix: Runners MUST use hosted LLM for first-pass generation

Skipping fresh-eyes

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

Not checking for identical implementations

  • Problem: Wasted compute on duplicates
  • Fix: Diff implementations before fresh-eyes

Forgetting cleanup

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

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.