skill-creator

The skill-creator skill helps users create new skills, modify and improve existing ones, and measure skill performance. It is used to build a skill from scratch, edit or optimize an existing skill, run evals to test a skill, benchmark performance with variance analysis, or optimize a skill's description for better triggering accuracy. It frames skill development as an iterative loop and helps figure out where the user is in that process before jumping in.

The high-level process it supports is: decide what the skill should do, write a draft, create a few test prompts and run Claude-with-the-skill on them, evaluate results both qualitatively and quantitatively, rewrite based on feedback, and repeat, then expand the test set for larger-scale runs. It includes scripts such as init_skill.py, run_eval.py, run_loop.py, aggregate_benchmark.py, improve_description.py, package_skill.py, quick_validate.py, and generate_report.py, plus an eval-viewer/generate_review.py for showing results and agents for analysis, grading, and comparison. A dedicated description improver script optimizes skill triggering. The skill also documents a DeerFlow sandboxed environment mode: when the skill_manage tool is available, all skill file operations (create, edit, patch, delete, write_file, remove_file) must go through skill_manage rather than sandbox write_file, because sandbox writes land in a per-thread output directory invisible to new chats. In DeerFlow, the package_skill.py and present_files steps are skipped since skill_manage persists skills to the per-user directory automatically, though eval workspace files can still use sandbox write_file.

It is aimed at users across a wide range of technical familiarity, and the skill advises adapting terminology accordingly, only using terms like JSON and assertion when the user shows they understand them and briefly explaining terms when in doubt. During intent capture it asks what the skill should enable, when it should trigger, the expected output format, and whether to set up test cases, suggesting test cases for objectively verifiable outputs and noting they are often unnecessary for subjective ones like writing style or art. The workflow remains flexible so users can skip formal evaluation if they prefer.