skill-creator

To document a Claude skill, you capture the user intent and define a clear skill objective, triggers, and success criteria. This process outlines scenarios and evaluation prompts to verify triggering and performance across edge cases.

Structuring a SKILL.md file involves crafting frontmatter and detailed instructions to load into context. The best way to structure it is by specifying dependencies, inputs, outputs, and constraints to effectively guide implementation and testing.

You run evals to test skill triggering by using eval-viewer and run_eval orchestration. This evaluates performance across edge cases and establishes a baseline for iterative improvement of your skill description and tests.

Yes, you need to specify dependencies, inputs, outputs, and constraints before drafting tests. Gathering these requirements and edge cases during the interview and research phase guides both implementation and testing.

Yes, you can optimize a skill description separately after running initial evaluations. You can optionally run a separate workflow to optimize triggering based on the feedback and benchmarking results from your eval baseline.

A skill fails to trigger correctly when the description and success criteria are not properly aligned. You need to refine the description and tests based on evaluation feedback and benchmarking to improve triggering accuracy.