chief-of-staff

AI Chief of Staff

A personal Chief of Staff that runs on top of an agent runtime (Claude Cowork / Codex): scheduled and on-demand skills, backed by a plain-text Markdown memory store, that prepares you for what's ahead, closes loops, coaches you, and gets sharper the longer it runs. It installs and configures itself by interviewing you.

It does not act on the world by default. It proposes; you approve.

Build artifact of [[AI Chief of Staff — Build Plan]]. Design: [[AI Chief of Staff]] · validation: [[AI Chief of Staff — Architecture Evaluation]] · loop spec: [[AI Chief of Staff — write-back method]].

Install

As a plugin (easiest) — Claude Code:

/plugin marketplace add fontesgerards/chief-of-staff
/plugin install chief-of-staff@chief-of-staff

Then, in the folder where you want your brain to live: /cos-onboarding. Codex uses the same engine/ package — see PUBLISH.md. Publishing steps + the naming/collision note: PUBLISH.md.

Or clone (dev / no plugin): git clone …, ./setup.sh, open the folder, /cos-onboarding. See INSTALL.md.

The engine is the shared plugin; your instance (private brain) is created by /cos-onboarding in your own folder — it never lives in the plugin.

The two halves

chief-of-staff/
├─ CLAUDE.md / AGENTS.md  ← always-loaded entry (Claude / Codex) → engine/INSTRUCTIONS.md
├─ setup.sh / INSTALL.md  ← one command to make skills invocable (/cos-onboarding, $cos-onboarding)
├─ engine/      ← git, shareable, identical for everyone (THE PRODUCT — no personal data)
│  ├─ INSTRUCTIONS.md   global behavior + autonomy + memory-access conventions
│  ├─ skills/cos-<name>/SKILL.md   invocable skills — cos-onboarding, cos-meeting-prep, cos-consolidate-memory, …
│  ├─ methods/          write-back, problem-solving, minto, coaching, PEI
│  ├─ templates/        entity schemas, memory-file + capture + proposal templates
│  └─ docs/             capability spike + write-isolation config + build notes
├─ .claude/skills → engine/skills   (symlink — Claude discovers /<name>)
├─ .agents/skills → engine/skills   (symlink — Codex discovers $<name>)
└─ instance/   ← local + private, GITIGNORED here; its own separate backup repo (one person's brain)
   ├─ config.md         identity, goals, autonomy level, connectors, schedules
   ├─ memory/           core · semantic · episodic · procedural · sources · archive
   ├─ state/            current run, open loops, commitments, pending questions, corrections
   ├─ queue/            proposals awaiting approval (outbound · approvals · review)
   ├─ log/              run history + maintenance changelogs
   └─ index/            optional, generated, rebuildable from Markdown

Why the split: to productize later, ship engine/ and let the onboarding skill regenerate a fresh instance/ for each new person. Nothing personal ever lives in the shared repo.

Status

v1 scaffold complete; skills are invocable as commands (run ./setup.sh, then /cos-onboarding in Claude or $cos-onboarding in Codex — see INSTALL.md). U0 capability spike resolved (2026-06-04): scheduling ✓, connectors ✓, write-isolation ✓ (structural — enforced per-run via Claude Code permissions.deny+sandbox or Codex permissions profile; recipes in engine/docs/write-isolation-config.md). KTD-5 stays "structural." Next: run /cos-onboarding to seed a real instance/, and add a private remote to the instance backup repo.

Core principles

1. Engine vs. instance are separate — the shareable framework never contains anyone's personal data.
2. Plain text is the brain — memory, methods, and skills are Markdown; the runtime is swappable.
3. Propose, never act (by default) — every outward action lands in a review queue. Autonomy is a dial, not a hardcode.
4. Inward writes ≠ outward actions — memory writes are made safe by append-only capture + git-reversible consolidation + confidence tiers, not by gating every write on approval.
5. Observe first, ask by exception — learn how you work from email/docs/calendar; only ask where genuinely uncertain.
6. It must maintain itself — memory and the system both decay; scheduled maintenance is first-class.