Skip to content
IonDenPublic

About

Agent skills for Claude Code and OpenAI Codex: audit subagents (subagent-optimizer), shrink SKILL.md files (skill-optimizer), and write tests that can fail (writing-tests-that-can-fail).

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

Agent skills for Claude Code and OpenAI Codex

validate License: MIT skills.sh Claude plugin directory: listed

A small catalogue of agent skills I use day to day, for people who run Claude Code or Codex and want to spend fewer tokens. Two skills make your agents and skills cheaper to load; the third keeps the tests your agent writes honest. Each skill is a folder with a SKILL.md. A skill lands here when I could not find a good public equivalent.

Skills

Skill What it does
subagent-optimizer For agent definitions: Claude Code subagents (.claude/agents/*.md) and Codex custom agents (.codex/agents/*.toml). Audits them and cuts what each agent costs to launch: a tight tools allowlist for Claude agents, no keys that make Codex skip an agent, less prompt bloat, a model and reasoning effort that fit the job. One real agent went from 20,122 to 13,962 launch tokens with the same result.
skill-optimizer For agent skills, the folders with a SKILL.md. Cuts a skill's SKILL.md down without losing an instruction, so it costs fewer tokens each time it loads: freezes every rule, anchor and literal in the original, then gates a rewrite against that freeze and rejects or asks rather than guessing. On five real skills it cut 1.2% to 3.5% of the body, 2.3% overall, kept every requirement it extracted, and listed the cuts it was unsure of instead of making them (evidence).
writing-tests-that-can-fail For writing, reviewing and fixing tests in any language. Keeps tests real: they exercise real behaviour rather than mocks, cover what matters and every condition in it, fail on the bug they exist to catch, and stay untouched when the code is what's wrong. Given the same prompts, Claude Haiku without the skill mocked both collaborators in 3 of 3 Python suites and copied expected values from the code's constants in 2 of 3; with the skill, 0 of 3 did either (evidence). In our Claude Code tests, two sentences in CLAUDE.md made it load far more often when tests were written (how).

Install

Any agent, with the skills CLI. The skills are also listed at skills.sh/ionden/skills:

npx skills add IonDen/skills                      # pick skills interactively
npx skills add IonDen/skills --skill subagent-optimizer -g -a claude-code -y
npx skills add IonDen/skills --skill subagent-optimizer -g -a codex -y
npx skills add IonDen/skills --skill skill-optimizer -g -a claude-code -y
npx skills add IonDen/skills --skill skill-optimizer -g -a codex -y
npx skills add IonDen/skills --skill writing-tests-that-can-fail -g -a claude-code -y
npx skills add IonDen/skills --skill writing-tests-that-can-fail -g -a codex -y
npx skills update

Most agents share .agents/skills, so the CLI installs there by default. Claude Code reads .claude/skills instead, which is why -a claude-code is on the lines above: without it the CLI asks which agents you want and Claude Code is not pre-selected. Pass --copy if you would rather have real files than symlinks.

Claude's plugin directory: the plugin is listed there. Add it on claude.ai, and it becomes available in your Claude Code sessions.

Claude Code, as a plugin:

/plugin marketplace add IonDen/skills
/plugin install ionden-skills@ionden

Codex, as a plugin:

codex plugin marketplace add IonDen/skills
codex plugin add ionden-skills@ionden

Development

Maintaining a plugin release? See publishing instructions for the upload ZIP and the OpenAI directory steps.

python3 -m pytest tests -q

The skills' own scripts are plain Python 3.10+ with no dependencies. The repository validator and the tests need pyyaml and pytest. Every test names the one-line bug that would make it fail.

License

MIT.

About

Agent skills for Claude Code and OpenAI Codex: audit subagents (subagent-optimizer), shrink SKILL.md files (skill-optimizer), and write tests that can fail (writing-tests-that-can-fail).

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages