Skills¶
Coding-agent skills for Claude Code and Codex, installed with the skills CLI.
What each skill changes¶
Every skill that ships evaluation cases is measured the same way: the same task is run with the skill available and without it, and a judge grades both against the same written assertions. The columns below are the share of assertions that passed in each arm.
A skill with no measurable difference is not a broken skill. It means the evaluated model already does that thing unprompted.
| Skill | With | Without | Difference |
|---|---|---|---|
codex-worker-prompting |
— | — | not measured |
colab-uv-training |
— | — | not measured |
github-cgd-identity |
— | — | not measured |
github-comment-attach-files |
— | — | not measured |
herdr-chezmoi-apply |
— | — | not measured |
herdr-meta-orchestrate |
— | — | not measured |
herdr-orchestrate-workers |
— | — | not measured |
herdr-tab-status |
— | — | not measured |
manage-agent-guidance |
— | — | not measured |
manage-public-private-dotfiles |
— | — | not measured |
manage-public-private-skills |
100% | 67% | +33 points |
python-transformers-convert |
— | — | not measured |
python-uv-workflow |
— | — | not measured |
research-before-implementation |
— | — | not measured |
research-high-impact-journal-publishing |
— | — | not measured |
shellscript-shdoc-docs |
— | — | not measured |
writing-kosshi |
95% | 66% | +29 points |
writing-router |
— | — | not measured |
writing-technical-slides |
— | — | not measured |
writing-telegraph |
— | — | not measured |
2 of 20 skills carry a recorded evaluation.