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

any-percent

Skill1 starsChanged 4 months agoMITView on GitHub

Runs 2-5 parallel agents in isolated git worktrees, each implementing a distinct architectural approach via Cerebras hosted LLM, then judges all survivors to select the winner (~60% token savings vs Claude direct). Use when the best architecture is unclear and you want real competing implementations to reveal the answer rather than deliberating upfront.

The file

We split this skill into the rules it gives an agent

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

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    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:

    Core principle: When unsure of the best architecture, implement multiple approaches fast via hosted LLM and let real code + tests determine …

    line 12
  2. Distinctive:

    Do NOT brainstorm extensively.

    line 53
  3. Distinctive:

    Variant limits: Max 5-6: Don't do full combinatorial explosion:

    line 85
  4. Distinctive:

    Create variants along that axis

    line 87
  5. Distinctive:

    Step 1: Gate check: - All tests pass

    line 118
  6. Very specific:

    Use scenario-testing skill.

    line 122

≈ “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

Any%

Explore different approaches in parallel - no restrictions, fastest path to comparing real implementations. Each variant uses hosted LLM for code generation.

Announce: "I'm using speed-run:any% for parallel exploration via hosted LLM."

Core principle: When unsure of the best architecture, implement multiple approaches fast via hosted LLM and let real code + tests determine the winner.

When to Use

  • Unsure which architectural approach is best
  • Want to compare 2-5 different designs quickly
  • "Try both approaches", "explore both", "not sure which way"
  • Want token-efficient parallel exploration

Workflow Overview

PhaseDescription
1. ContextQuick gathering (1-2 questions max)
2. ApproachesGenerate 2-5 architectural approaches
3. PlanCreate implementation plan per variant
4. ImplementDispatch ALL agents in SINGLE message, each uses hosted LLM
5. EvaluateScenario tests -> fresh-eyes -> judge survivors
6. CompleteFinish winner, cleanup losers

Directory Structure

docs/plans/<feature>/
  design.md                  # Shared context
  speed-run/
    any-percent/
      variant-<slug>/
        plan.md              # Implementation plan for this variant
      result.md              # Final report

.worktrees/
  speed-run-variant-<slug>/  # Variant worktree

Phase 1: Quick Context

Gather just enough to generate approaches (1-2 questions max):

  • What are you building?
  • Any hard constraints? (language, framework, infrastructure)

Do NOT brainstorm extensively. The point of any% is to explore fast, not deliberate slowly.

Phase 2: Generate Approaches

Identify the primary architectural axis (biggest impact decision):

  • Storage engine (SQL vs NoSQL vs file-based)
  • Framework (Express vs Fastify vs Hono)
  • Architecture pattern (monolith vs microservices vs serverless)
  • State management approach
  • Auth strategy

Generate 2-5 approaches along that axis:

Based on the requirements, here are 3 approaches to explore:

1. variant-sqlite: SQLite storage with query builder
   → Pros: Simple, embedded, zero config
   → Cons: Single writer, no replication

2. variant-postgres: PostgreSQL with ORM
   → Pros: Scalable, ACID, rich queries
   → Cons: External dependency, more setup

3. variant-redis: Redis with persistence
   → Pros: Fast, built-in pub/sub
   → Cons: Memory-bound, limited queries

All will be implemented via hosted LLM and tested.
Spawning 3 parallel variants now.

Variant limits: Max 5-6. Don't do full combinatorial explosion:

  1. Identify the primary axis
  2. Create variants along that axis
  3. Fill secondary slots with natural pairings

Phase 3: Plan + Implement (Parallel)

Setup worktrees:

.worktrees/speed-run-variant-sqlite/
.worktrees/speed-run-variant-postgres/
.worktrees/speed-run-variant-redis/

Branches:
<feature>/speed-run/variant-sqlite
<feature>/speed-run/variant-postgres
<feature>/speed-run/variant-redis

CRITICAL: Dispatch ALL variants in a SINGLE message

<single message>
  Task(variant-sqlite, run_in_background: true)
  Task(variant-postgres, run_in_background: true)
  Task(variant-redis, run_in_background: true)
</single message>

Variant agent prompt: see references/variant-agent-prompt.md

Phase 4: Evaluate

Step 1: Gate check - All tests pass

Step 2: Run same scenario tests against all variants

Use scenario-testing skill. Same scenarios, different implementations.

Step 3: Fresh-eyes on survivors

Fresh-eyes review of variant-sqlite (N files)...
Fresh-eyes review of variant-postgres (N files)...

Step 4: Invoke Judge Skill

CRITICAL: Invoke speed-run:judge now. See references/judge-invocation.md for full context and any%-specific scoring notes.

Phase 5: Completion

Winner: Use finish-branch

Losers: Cleanup

git worktree remove .worktrees/speed-run-variant-sqlite
git worktree remove .worktrees/speed-run-variant-redis
git branch -D <feature>/speed-run/variant-sqlite
git branch -D <feature>/speed-run/variant-redis

Write result.md:

# Any% Results: <feature>

## Approaches Explored
| Variant | Approach | Tests | Scenarios | Fresh-Eyes | LLM Calls | Result |
|---------|----------|-------|-----------|------------|-----------|--------|
| variant-sqlite | Embedded SQL | 18/18 | 5/5 | 0 issues | 3 | WINNER |
| variant-postgres | External DB + ORM | 20/20 | 5/5 | 1 minor | 4 | eliminated |
| variant-redis | In-memory + persist | 16/16 | 4/5 | 2 issues | 5 | eliminated |

## Winner Selection
Reason: Simplest architecture, zero external dependencies, all scenarios pass

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

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

Skill Dependencies

DependencyUsage
superpowers:writing-plansGenerate implementation plan per variant
superpowers:using-git-worktreesCreate isolated worktree per variant
superpowers:dispatching-parallel-agentsDispatch all variant agents in parallel
scenario-testing:scenario-testingRun same scenarios against all variants
fresh-eyes-review:fresh-eyes-reviewQuality review on survivors
speed-run:judgeScoring framework (bundled)
superpowers:finishing-a-development-branchHandle winner, cleanup losers

Common Mistakes

Too many variants

  • Problem: Combinatorial explosion, wasted compute
  • Fix: Cap at 5-6, pick primary axis only

Extensive brainstorming before exploring

  • Problem: Defeats the purpose of any% (fast exploration)
  • Fix: 1-2 questions max, then generate and go

Using Claude direct for all code generation

  • Problem: No token savings, defeats speed-run purpose
  • Fix: Variants MUST use hosted LLM for first-pass generation

Dispatching variants in separate messages

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

Skipping scenario tests

  • Problem: Can't compare variants fairly
  • Fix: Same scenarios against all variants

Forgetting cleanup

  • Problem: Orphaned worktrees
  • 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.