Skip to content

An agent organization structured as a company — 15+ departments, 125+ skills, each independently installable, citing the standards and regulators that settle the question. Runs in Claude Code and ChatGPT.

3.1 KB
Raw Blame

AGENTS.md

An agent organization structured as a company: 16 departments, 172 skills, each department installable on its own.

This file describes the repository to an agent working on it. The skills under plugins/ are a different thing — they describe how to do the work of a function, and they load when a request matches their description.

Installing, in either tool

The skills are identical. Only the manifests differ, and both are generated from the same tree, so a skill fixed once is fixed for both.

Manifests
Claude Code plugins/<department>/.claude-plugin/plugin.json, .claude-plugin/marketplace.json
ChatGPT / Codex plugins/<department>/.codex-plugin/plugin.json, .agents/plugins/marketplace.json

Before you open a pull request

./scripts/check-all.sh

Every check CI runs, in one script — the workflow calls this same file, so the two cannot drift. Stage new files first: the surface guard reads git ls-files, so an unstaged file is invisible to it and the check passes locally then fails in CI.

Rules that are enforced rather than suggested

  • One owner per path. docs/AGENT-SURFACES.md maps every tracked path to exactly one agent, and no two owners may claim the same path. A new department adds its roster row, its surface block and its charter in the same change.
  • Generated files are never hand-edited. The README, the org chart, the social card, each skill's references/sources.md, and these manifests are all emitted from the tree and verified with --check. An edit to one is lost on the next regenerate.
  • Sources are references, never copies. sources/*.toml maps outside authorities to the skills they settle questions for, and scripts/build-sources.py emits each skill's list into its own references/sources.md. The license field decides whether an agent may quote a source or only cite it — most of what a professional must cite is not open, so when torn between two classes take the more restrictive one.
  • Verticals emit one way. verticals/<slug>/ plus the core emits a standalone repository via scripts/build-vertical.py. Generated output is never hand-edited and is not committed; --verify emits to a temporary directory and runs the emitted repository's own checks instead.
  • US English, by exact word form. scripts/check-us-english.py --fix rewrites. It leaves URLs alone, which it did not always do.
  • No third-party licensed content. All prose here is original; scripts/check-provenance.py is the backstop.

Departments

  • executive — Office of the CEO
  • technology — Technology
  • it-operations — IT Operations
  • product — Product
  • marketing — Marketing
  • demand-generation — Demand Generation
  • revenue — Revenue
  • finance — Finance
  • operations — Operations
  • pmo — Program Management Office
  • people — People
  • legal-risk — Legal & Risk
  • security — Security
  • customer-experience — Customer Experience
  • data-analytics — Data & Analytics
  • corporate-strategy — Corporate Strategy