skill-creator

# Skill Creator

Create new skills, improve existing skills, and measure skill performance through a repeatable workflow.

## When to use

Use this skill when a user wants to create a skill from scratch, edit or optimize an existing skill, run evaluations, benchmark skill performance, improve trigger accuracy, or package a finished skill for distribution.

## Core workflow

1. Clarify the intended capability, trigger context, output format, dependencies, edge cases, and success criteria.
2. Write a first SKILL.md draft with focused frontmatter and progressive-disclosure instructions.
3. Create realistic positive and negative evaluation prompts.
4. Run with-skill and baseline comparisons in the same iteration.
5. Review qualitative outputs and quantitative assertions.
6. Read user feedback, identify repeated failure patterns, and revise the skill without overfitting to a single example.
7. Expand the evaluation set, repeat the benchmark, optimize the description, and package the final skill.

## Design principles

Keep SKILL.md concise, explain why important steps matter, prefer reusable workflows over brittle examples, and place trigger information in the description. Bundle deterministic helper scripts when several test cases would otherwise recreate the same work.

## Evaluation and review

Store evaluation prompts in evals/evals.json. Organize outputs by iteration and evaluation name. Record timing and grading metadata, aggregate benchmark results, and present the qualitative outputs together with the quantitative comparison. Stop when the user is satisfied, feedback is empty, or further changes no longer create meaningful improvement.

## Output contract

Return a clear stage assessment, the next recommended action, the files changed, the evaluation evidence, and any remaining risks or follow-up steps.