SKILLs: Instructions file for Codex

AGENTS.md

Project-agnostic library of agent skills: each SKILL.md is a procedure an LLM agent loads and follows. Downstream projects consume it as git submodule at .github/skills, as Claude Code plugin, or through npx skills, so every skill stays neutral across unrelated projects. Load a linked skill only when a task needs it.

Repository shape

• One directory per skill, kebab-case, containing file named exactly SKILL.md. Directory name MUST match skill's name in its frontmatter.
• Every top-level directory that is neither dotted nor skills/ IS a skill directory; that namespace is reserved. Repository plumbing lives in dotted directories (.github, .claude, .claude-plugin, .agents) or single files at root (e.g. sync-skills.example.yml).
• skills/ is vendor-neutral hub: one committed relative symlink per skill, skills/<skill-name> to ../<skill-name>. Every vendor path is one symlink to that hub, so another agent costs one link. .github/skills serves GitHub Copilot and mirrors path downstream projects mount this repository at; .claude/skills serves Claude Code; .agents/skills serves OpenAI Codex.
• .claude-plugin/ carries installer manifests: marketplace.json lists every skill by canonical root path, the address installers resolve; plugin.json declares same root container to Claude Code. Gate regenerates that list from tree, so adding a skill never edits a manifest by hand.
• LICENSE: MIT.
• Bundled Python is one uv workspace: root pyproject.toml lists every member; uv.lock pins the library.

Each skill loads its reference files on demand from references/.

Current skills:

• advisor/ - Reads a research project like a principal investigator: what it combines, what is outdated, what your lab can afford, cheapest experiment settling each claim.
• aesthete/ - Designs, builds, reviews web interfaces meeting WCAG 2.2, honoring supplied brand or design system, avoiding templated look of generated UI.
• asd-ste100/ - Writes, rewrites, checks text in ASD-STE100 Simplified Technical English, naming each unapproved word with approved alternatives.
• author-skill/ - Turns task history, workflow, or procedure into reusable SKILL.md another agent can follow with no memory of original session.
• caveman/ - Compresses replies into terse phrasing keeping every technical fact; can rewrite a prose file in place.
• draft-paper/ - Drafts conference, workshop, journal, survey, or demo papers and answers reviews, tracing each empirical claim to an artifact and each citation to a retrieved record.
• fact-check/ - Checks a document claim by claim against retrieved sources, quotes behind every verdict; changes nothing until you approve each correction.
• finance/ - Keeps a private record of slow-money stock bets with odds and kill dates, measured by deflated Sharpe, PBO, purged walk-forward at book's cap, sized by evidence under a ruin bound; operator replies in chat and trades by hand.
• git-commit/ - Drafts and reviews Conventional Commits messages; can commit and push in one step.
• humanize/ - Rewrites AI-sounding prose so it reads like its writer, keeping every claim's original strength.
• lit-review/ - Produces literature review where every citation traces to a paper retrieved from OpenAlex, arXiv, or Crossref, never memory.
• peer-review/ - Reviews a paper as an adverse referee would, pressing on claims, design, analysis, limitations, novelty.
• pl-theorist/ - Brings a programming-languages theorist's discipline to design, code, review, tests, tuned per language.
• ponder/ - Answers an open question and shows its work, sourcing every load-bearing claim, testing strongest rival explanation.
• ponytail/ - Forces laziest solution that works: standard library and native features before custom code or new dependencies.
• read-pdf/ - Extracts text and metadata from a PDF, answers questions about it with page-cited evidence, without OCR.
• reframe/ - Turns a planning discussion into testable direction judgment with costed routes and evidence that would prove it wrong.
• search-web/ - Searches web, Wikipedia, scholarly record when harness has no search tool; pulls readable text out of a page.
• setup-env/ - Installs a project's development toolchain in userspace: no sudo, no docker, nothing assumed present but uv.
• summon/ - Hands work to another agent: whether to delegate, what to tell the delegate, how to keep parallel agents apart, how to judge what comes back.
• thematic-analysis/ - Develops themes from qualitative text under one named school, each theme backed by verbatim extracts and counts.