CompanyOS, an operating system in markdown

A complete, opinionated, fork-friendly operating system for an AI-instructed company. Three layers, in order of reach:

  1. Skills. Markdown patterns an AI agent loads to inform judgment. Decision memos, pre-ship adversary attacks, diagnose-before-acting under pressure, vocabulary discipline, memory writing. Thirteen skills, all universal.
  2. A justfile template. A discoverable command surface for the human-or-agent runbook layer. Health checks, smoke tests, deploy dances, session-end rituals.
  3. Constitutional documents. Operator-owned strategy that agents read but never rewrite. An amendment-flow pattern keeps strategy and execution separate without bureaucracy.

The same 13 skills run inside the company that built this Tome. That's the dogfooding test: if a skill isn't useful enough for our own daily work, it's not useful enough to ship to anyone else.

Why this exists

Most "company operating manual" content is one of two things:

  • A blog post describing what one team does, not portable.
  • A framework abstraction with no implementation, not actionable.

This Tome is the third option: a forkable artefact. The skills are written so they work for a company that has nothing to do with the one that authored them. The justfile is a template. Your project's recipes live in your repo, not here. The constitutional pattern is a shape, not a content claim.

What you get

companyos/
├── README.md                # this file
├── AGENTS.md                # root operating context (load at agent start)
├── CLAUDE.md → AGENTS.md    # alias for Claude Code
├── tome.json                # manifest (Tome spec v1)
├── justfile.template        # generic operator runbook starter
├── skills/                  # 13 markdown skills
│   ├── decision-memo.md
│   ├── pre-ship-adversary.md
│   ├── adversarial-test.md
│   ├── diagnose-before-acting.md
│   ├── vocabulary-check.md
│   ├── hard-evidence-under-pushback.md
│   ├── strategic-realignment.md
│   ├── memory-write.md
│   ├── incident-recovery.md
│   ├── multi-target-publish.md
│   ├── first-time-action-gate.md
│   ├── changelog.md
│   └── session-debrief.md
└── docs/
    ├── three-layer-pattern.md      # how skills + justfile + constitutions compose
    └── memory-taxonomy.md          # the 4-type memory pattern

How to fork

  1. Copy the skills/ directory into your project's agent-skills location (.claude/skills/, .cursor/rules/, .github/copilot-instructions/skills/, depending on your tools; every major AI dev tool has its own location).
  2. Adopt the AGENTS.md (or rename it to your tool's preferred filename, such as CLAUDE.md or .cursorrules). It's intentionally minimal; layer your project specifics on top, don't replace it.
  3. Copy justfile.template to your repo's justfile and replace the placeholder recipes with your project's actual commands. The shape (groups, conventions, idempotency norms) is the contribution; the recipes are yours.
  4. Read docs/three-layer-pattern.md once. The architectural decisions (why skills are markdown, why constitutions are operator-owned, why memory is separate from skills) inform everything.

That's it. The skills will start firing on the first session.

What we kept private and why

This Tome is the patterns. The implementation behind it (the company's actual outreach genome, our specific search index schema, our growth-engine A/B logs, our memory contents) stays private because it's load-bearing competitive work. The pattern is the gift; our specific implementation is ours.

The honest line that goes here:

"Buffer's open salaries didn't include their lead-scoring algorithm; Linear's changelog doesn't expose their pricing model decisions. We share the operating system, not the secret sauce."

If you fork this and want to share your operating system back, we'd love to see it. Open a PR or fork it under your own org. The shape generalises faster when more shapes exist.

What's not here yet

  • Project-specific runbooks (your company's freeze-recovery script, your deploy dance). The pattern is in incident-recovery.md and multi-target-publish.md; the recipes are in your justfile.
  • A web UI for browsing skills. The Tome is intentionally markdown-first; surface it however you like.
  • Quality / security scoring rules. Different companies have different threat models; pick your own.

Versioning

Manifest at tome.json follows the Tome spec v1. Skill IDs are stable; descriptions and bodies evolve. Material breaking changes (rename, remove, shift in semantics) trigger a major version bump.

CompanyOS is a Tome from TomeVault, the neutral referee that keeps AI agents behaving the way you taught them as the models underneath keep changing. We grade the files that steer your agents on whether they load, stay clear, run safe, and hold up across models, then keep that verdict true as the tools change. Scan your own instruction files and read the verdict at tomevault.io/scan.

Maintained by: TomeVault. License: MIT. Fork freely. Filing issues: GitHub issues on this repo. Follow updates: the Tome page on TomeVault for change-log + version history.