Skill Creator

Create skills that supply non-obvious guidance for a defined task. Preserve the user's scope, existing authorization, invocation metadata, and licenses. Assume the agent can handle ordinary programming and tool use.

Design

• Describe the capability and its actual trigger briefly. Avoid lists of every supported feature or broad keywords that attract unrelated tasks.
• Keep the root self-contained for simple work. For substantial modes, link supporting references with a clear read condition. Do not load all modes.
• Describe outcomes and decision criteria for flexible work. Reserve fixed sequences for fragile operations or concrete correctness constraints.
• Add scripts when repeated mechanics justify them, references for conditional detail, and assets for reusable output. Inspect callers before removing existing resources. Do not add empty scaffolding or duplicate documentation.
• Keep existing invocation policies. For new skills, allow automatic discovery unless the user requests explicit-only invocation. Sensitivity alone is not a discovery gate; check authorization before the consequential action.

Create or Update

Use the request and existing package to establish the trigger, output, and constraints. Ask only when missing information materially affects the result. A narrow update may need only an edit and focused validation, not initialization or package planning. Do not turn a past example into a universal requirement.

For new packages, use scripts/init_skill.py when it helps. Preserve a requested location and create only resources with a current use. Use scripts/package_skill.py only when an archive is part of the requested output.

References

• Read design-principles.md for context placement and degree-of-freedom decisions.
• Read package-layout.md when choosing package files and resource types.
• Read creation-workflow.md for a new or substantially revised package, adapting its steps to the task.
• Read workflows.md or output-patterns.md when a workflow or output contract needs examples.

Validation

Run scripts/quick_validate.py <skill-dir>. Check reference reachability, metadata preservation, and trigger precision. Run new or changed scripts against representative inputs. Structural validation does not prove useful behavior.

For a complex or risky skill, use an independent agent with a realistic request and raw artifacts when delegation is available. Do not supply the intended answer. Keep evaluation side effects within the authorized scope, inspect the result, and revise only where observed behavior warrants it.