Describe the job.
Write the skill’s purpose, prerequisites, and expected behavior. Keep the trigger narrow enough to leave unrelated work alone.
The local workbench / Build. Compare. Refine.
Skill Studio
Write what good looks like. Test whether your skill gets you there. Build focused expertise with evidence you can inspect.
v1.0.63 · signed & notarized · all platforms, checksums & install notes →
Or open it from your terminal
Local workspace. Your configured model provider handles model requests.
Make the improvement observable
From a useful idea
to a tested skill.
Start with a real task. Describe the expected result, compare configurations, and inspect failures before you publish. Shorter instructions can be a better outcome.
Change one thing.
Hold the task steady. Compare skill versions or configurations, then inspect what succeeded, failed, or became slower.
Keep the evidence.
Use results to improve or retire instructions. Recheck after meaningful model, harness, or tool changes.
Inside the workbench
Designed for iteration.
Your skill, test cases, and results belong together. Studio helps you work through them without turning every comparison into a custom script.
Plain-English Evals
Define expected behavior in your own words. Inspect the judge’s reasoning and check that the verdict matches the outcome you care about.
A/B Compare
Compare prompts and skill configurations side by side. Use the same task and starting conditions so the difference is meaningful.
Your Model
Use supported providers and your configured models. Re-evaluate when the model, tools, or harness changes; results do not transfer automatically.
Local Workspace
Author and manage skills on your machine. Cloud model calls send evaluation inputs to your selected provider; publishing sends content to its destination.
Model ≠ agent
Test the setup
you actually use.
A model operates inside a harness with its own instructions, tools, and permissions. A skill that helps one setup may add overhead to another.
Read the evaluation guide ↗- Hold the task steady
- Same request, same starting files, same pass conditions.
- Record the configuration
- Skill version, model, provider, harness, tools, and effort where available.
- Interpret within scope
- A result describes the tested setup. It is not a universal model ranking or a safety certification.
Start with one task
Get Started
Use the desktop app or launch the same workspace from your terminal. Connect the model provider you want to evaluate with.
Run in your project directory
First time? Run npx vskill@latest init to scaffold your test suite. Read the Studio guide ↗
Get the desktop app ↗Native app · v1.0.63