Create, run, rate, and iterate on your Claude Skills.

This is alpha software written for personal use.

• It runs Claude in yolo mode which can and will wipe your data.
• It can also burn through a ton of tokens if your Skills aren't lean.
• It is 100% vibecoded, and this time I have not read the code.

If you end up burning all your tokens only to brick your computer, I'm not responsible.

Create a new workspace:

npx woodshed create my-idea

That gives you a place work on your Skills:

cd my-idea

Time to shed:

npx woodshed

By default, the entire matrix runs 10 times.

Shed is idempotent so by default re-running it will instantly "skip over" past results as if they happened instantly.

Pass npx woodshed --reset to force re-runs. This will delete each re-run's data in results/ before attempting it. You can also delete the results/ yourself if you want.

Take a close look at the results/ folder after your first successful run. It contains the log from the main agent with your fixture's prompt.md, the log from the evaluating agent with your fixture's eval.md, the workspace folder in which they both run, and probably some other junk.

my-idea/
  skills/
    # Skills you want to create or refine
    my-skill/
      # Each Skill can have one or more variants
      baseline/SKILL.md
      experiment/SKILL.md
      silly/SKILL.md
  fixtures/
    # Fixtures that test your Skill(s)
    my-fixture/
      prompt.md
      eval.md
  assets/
    # Data shared by fixtures
    words.txt
  results/
    # Outputs and past runs appear here

I recommend running another Claude instance and to talk with it about the /results.

Then you can use insights from that convo to refine your eval.md and SKILL.md.

Tip: If you're iterating on a skill, ask Claude to write down each experiment in a doc so you can see what works and what doesn't.

--runs <n>      Number of runs per variant (default: 10)
--reset         Delete old results and start fresh
--reeval        Re-run evaluation on existing workspaces
--cache-only    Show cached results only

MIT