Skill Creation and Iteration Assistant

Problem context

Turning a workflow into a reusable skill is harder than drafting SKILL.md. It is easy to write a plausible description, but harder to know whether the skill triggers in the right situations, produces the expected output, and improves over the no-skill baseline. Relying on a few informal runs can bias the feedback, especially after multiple revisions.

How it works

The skill guides a practical iteration loop: clarify intent, triggers, inputs, outputs, and acceptance criteria; draft SKILL.md, with special attention to a trigger-oriented description; then create 2–3 realistic test prompts. It runs with-skill and baseline executions side by side, drafts quantitative assertions while runs are in progress, grades outputs against the assertions, and aggregates pass rate, duration, and token usage into a benchmark view. For more complex work, it supports a separated planning, generation, and evaluation flow with handoff notes, making each revision easier to compare.

Boundaries

It is best when the skill needs verification, comparison, and repeated refinement. For quick subjective style work, the full eval loop can be skipped and the author can iterate directly on feedback. The source also cautions against creating misleading, malicious, or unauthorized-access capabilities.