Skill Creator

Problem

Agent skills can easily become a pile of prompts without clear triggers, output formats, or verifiable results. If the SKILL.md description is vague or the boundaries are unclear, the model may fail to use the skill when it is useful, or force subjective tasks into a rigid workflow.

How It Works

Skill Creator follows a capture–draft–test–evaluate–iterate loop:

• Capture intent: confirm what the skill should do, when it should trigger, the expected output, and whether evals/evals.json test cases are needed.
• Draft the skill: fill in name, description, compatibility, and the body; the description drives triggering and should cover synonymous user contexts.
• Run tests: launch with-skill and baseline subagents in the same turn, saving results under directories such as iteration-1/.
• Quantitative evaluation: draft assertions while runs are in progress, then grade, aggregate benchmark.json, and open a viewer for qualitative outputs and metrics.

Boundaries

Best for skills with objective outputs or fixed workflows, such as file transforms, data extraction, or code generation. Writing style or design work is better judged qualitatively. Create a new directory for each iteration to preserve history and compare baseline and skill versions. Assertions should be objectively verifiable, avoiding hard metrics for subjective preferences. Do not use it to create misleading, unauthorized-access, or malicious skills.