Skills for Claude Code and Codex. The durable loops drive a repository toward a goal. The parallel skills split a review, an investigation, a research question or a change across independent agents. The rest cover shipping and growing a repository, replying, recovering cut-off work, and explaining.
/investigate,/review-pr,/review-code,/researchand/implementsplit the work across parallel agents, scaled to the size of the task, and check every finding before acting on it./sergio-loopruns a durable goal loop in one repository, one small slice at a time, with goals and evidence indocs/and resumable state in.omc/sergio-loop/.- An optional global Stop hook continues the loop in the same session and stays inert everywhere else.
/refine-loopmakes small, behavior-preserving improvements ranked byROI = Impact × Confidence ÷ Effort, one per round, each checked by a fresh reviewer./docs-loopaudits documentation against the default branch, as a dry run, with staged edits, or with one pull request per repository. It never merges.- None of them treats a stored goal as permission to merge, publish or deploy.
/kb-researchand/kb-reviewanswer and review from your team's own record, through one adapter file you fill in for whatever knowledge base you have./grow-my-repoplans where an open-source repository gets listed (GitHub itself, registries, directories, public catalogues, media, communities, funding) and does the submissions an agent may do./submit-awesomesends one entry per awesome list, by each list's rules. Both hand the maintainer a packet wherever a place wants a human./file-propens pull requests whose title says why the change matters and whose body starts with the problem./verified-replydrafts a reply you send as yourself, every claim checked against a live source first./revivebrings work cut off by a usage limit, a crash or a compaction back from where it really stopped, and checks the world before every write so nothing happens twice./eli5explains work or a system in plain words with real names, and/help-decidelays out a decision so one word answers it.- A validator checks every skill package, and checks each loop against the shared loop contract.
git clone https://github.com/GeiserX/agent-skills.git && cd agent-skills
mkdir -p ~/.claude/skills && ln -s "$PWD/skills/sergio-loop" ~/.claude/skills/sergio-loopThen run /sergio-loop <goal> in a fresh Claude Code session. If a skill of the same name is already installed, the installer moves it aside first and also installs the Stop hook. In Codex, link from codex/skills/ into ~/.agents/skills/ and call a skill as $<name>; the Stop hook is Claude Code only.
The full documentation is at geiserx.github.io/agent-skills.
- Getting started: the skill and hook installers,
settings.jsonentries, canonical sources and local overlays - Usage: what each skill does, its defaults and invocations, the goal loop, the parallel skills, the knowledge-base skills, and the skills for growing a repository, pull requests, replies and recovery
- Development: validating the skill packages, rebuilding the Codex copies, running the tests
- Related projects: the skills from other projects kept in
vendor/, and the repository this one replaced
vendor/ holds skills written by other people, kept here as pinned snapshots with their licences. They are not installed by default; link the ones you want the same way as the others.
- pstack by Lauren Tan (MIT), 52 skills such as
unslop,blast-radiusandshow-me-your-work, taken from Michael Denyer's Claude Code port. Sources, commits and local changes are in vendor/pstack/PROVENANCE.md. writing-for-agentsfrom Matt Pocock's skills (MIT). See vendor/mattpocock-skills/PROVENANCE.md.