Skill Studio

Describe expected behavior in plain English and create realistic test cases for the skill. Run them with your selected model, then inspect outputs and the judge’s reasoning. Include cases where the skill should stay out of the way. A passing judge verdict is evidence for that test configuration, not a guarantee of correctness.

Hold the task, starting files, and pass conditions steady. Compare a baseline with a skill, or compare two skill configurations. Review the outputs and judge reasoning alongside pass rates. Results describe the model, harness, tools, and settings you tested; they are not a universal model ranking.

Choose from the providers and models supported by your installed version. Configure credentials for cloud providers, or a supported local model runtime for local inference. Model availability and tool support vary. Re-run your evaluation after changing the model, provider, instructions, or tools.

Skill drafts, test cases, and evaluation results are managed in your local workspace. A cloud evaluation sends the prompt and relevant inputs to your selected model provider. Registry discovery and publishing also use the network. For local-only inference, select a supported local model and check the tools it can call. Local storage alone does not make a workflow air-gapped.

Review your repository, visibility, and files before using Publish. Studio can help create a GitHub repository and push the skill through your configured GitHub access. Public registry submission is a separate sharing destination. Keep private skills in their private repository and use authorized GitHub access to install them. A published repository or completed scan does not prove that a skill improves your task. See the submission guide and security guidelines.

Review the target agent and destination shown before installing. Project scope applies to the selected project. User scope applies to that user’s projects on the current machine. Global scope uses shared system paths on that machine and may require additional permissions; it does not synchronize other computers. Agent-specific paths and prerequisites vary. One supported installation target does not mean every harness executes the skill identically.

An update notification identifies a newer skill version. Review the changes and the selected installation destination before applying it, then check the resulting version. Re-run relevant evaluations after a meaningful update. Skill updates and desktop application updates are separate; desktop release details live on the download page.

• Security guidelines — the security-scan pipeline that gates every published version.
• Plugins — install Studio functionality as Claude Code plugins.
• Submitting skills — manual submission flow for authors not using Studio's Publish action.