How I rolled out Claude across a startup without fragmenting the company — personal accounts for execution, a governed layer around them, and a Chief-of-Staff agent for every leader.
Most companies are adopting AI exactly backwards. They buy seats, tell everyone to “use AI,” and a quarter later they’re faster and dumber: the best prompts, frameworks, and decisions live in ten thousand private chat histories that nobody else can find. Individual productivity went up. Institutional capability didn’t move.
I’ve started four companies and led growth at a few more, and the pattern is always the same — the tool isn’t the hard part. The operating system around the tool is. So I wrote one down, and then I built it with Claude as my co-author. This is the short version. The full, copy-paste-ready playbook is free at the bottom.
The bet: personal accounts for execution, a governed layer around them
Start with the money. The personal-vs-Team decision is, before anything else, a cost-containment call. Metered API usage for a few heavy, agentic users can balloon unpredictably; personal Claude accounts (Max for power users) turn that into a fixed per-seat line item — flat cost, no metered surprises — which is exactly what you want while a small team is still figuring out how it works. So I run personal accounts as the execution layer. The catch: personal accounts give you nothing for collaboration, governance, or memory.
So you build that layer yourself, on three pillars you already need:
- A system of record — one repo (I use GitHub) that is the canonical home for prompts, playbooks, templates, SOPs, your metric dictionary, and your agent configs.
- A secrets manager — one control plane (1Password) so credentials are retrieved at runtime, never pasted into a chat.
- An identity provider — Google Workspace for SSO and the join/leave lifecycle of every other tool.
Round it out with Linear for workstreams, Mixpanel + GA4 for instrumentation, and Slack for the “new canonical asset” announcements, and you have a stack where private AI output gets promoted into shared, governed, discoverable institutional memory. That promotion path — draft → review → merge → announce — is the single habit that separates “we use AI” from “we compound with AI.”
The real unlock: a Chief-of-Staff agent for every leader
Here’s where it stops being a knowledge-management story and becomes an operating system.
Each functional leader — CEO, Product, Marketing, Sales, CS, Finance — gets their own Chief-of-Staff (CoS) agent. It triages that leader’s workstreams, keeps their slice of the company’s source-of-truth current (pulling from Linear, auto-captured meeting notes, and analytics), and prepares their dashboard view and weekly digest. It proposes; the human decides. Nothing it touches ships without a human gate.
Around the CoS agents sit two adjudicators — read-only agents that are the single source of truth for their domain. An analytics-lens owns every metric definition and guards against vanity numbers; a finance-lens owns the money and judges whether any decision moves you toward or away from break-even. Before a recommendation reaches a human, it’s checked against both.
And optionally, a red-team agent anyone can switch on: a constructive devil’s advocate that surfaces your unstated assumptions, argues the strongest counter-case, and names the failure modes before you commit. The CoS proposes; the red-team pressure-tests; you decide. (For parallel build work, these same role definitions get spawned as a runtime agent team — a lead plus teammates with a shared task list — but most of the daily value is the single CoS.)
Pro tip: tell your CoS agent to be proactive — search first, answer, then ask
The biggest behavioral unlock isn’t delegation; it’s defaulting your CoS agent to find answers itself. A weak assistant pings you with “what’s our LinkedIn URL?” / “which quarter?” / “do we have a doc on this?” A strong one searches the repo, the web, the analytics, and the meeting notes, comes back with the answer (and its sources), and only escalates the genuine forks that are actually yours to call. Bake it into the role: “Be proactive. Research and answer before asking. Only surface decisions I alone can make.” It turns a chatbot you have to babysit into a chief of staff who clears your desk.
One north star, wired into everything
The connective tissue is a single Org North Star — for the product, not the rollout. One sentence of vision, one metric that best proxies delivered customer value, and a viability constraint (e.g. “grow that metric and move toward break-even”). It’s defined once and referenced everywhere, never redefined.
That’s what makes the system coherent instead of a pile of dashboards:
- The dashboard is designed around it — the CEO’s headline tile is the north-star metric, and each leader’s view is the one lever they own from the Pirate Metrics (AARRR) funnel — Acquisition, Activation, Retention, Revenue, Referral (Dave McClure’s framework).
- Every CoS agent weighs and ranks decisions by their expected effect on that metric, within the viability constraint. Anything that doesn’t move it gets deprioritized and flagged.
- The red-team’s first challenge to any plan is: does this actually move the north star, or just a vanity proxy? Show me the causal evidence.
When the metric, the OKRs, the agents, and the dashboard all point at the same place, the whole company reads from the same destination.
The honest trade-offs (read these before you copy anything)
This model is not free, and pretending otherwise is how rollouts fail:
- Data handling. Consumer/personal Claude plans run on different data terms than commercial ones. Decide up front what data classes may touch a personal account. Anything client-confidential or regulated belongs on a commercial (Team/Enterprise) surface — full stop.
- Account ownership. A reimbursed personal account belongs to the individual. You can’t centrally revoke it, and the history leaves when they do. Write the policy: corporate-email sign-up, output is company IP, explicit offboarding.
- It’s vendor-coupled by design. Keep your role specs and prompt packs in plain markdown so they’re portable, but know you’re optimizing for one provider.
There are clean ways to backfill what a Team plan would give you (central admin, search, audit) — and a clear set of triggers for when you should just upgrade. That’s all in the playbook.
Start absurdly small
You do not need the full architecture on day one. The Minimum Viable Rollout is one repo, one north star, one metric dictionary, one CoS agent (for the CEO or Head of Product), and one onboarding file. No dashboard, no agent teams yet. Time-to-first-value is days. Then you graduate through five phases — foundations, leadership pilot, a single-department end-to-end pilot, dashboard integration, and scaled rollout — only adding complexity once the prior step has earned it.
Take the whole thing
I wrote the full playbook to be executable, not inspirational: the operating model, the decision framework for personal-vs-Team, the agent configs, a role- and skill-aware onboarding skill you generate per user, acceptance criteria, and operating cadences. It’s one Markdown file, and it’s free under Creative Commons (CC BY 4.0 for the writing, MIT for the code/templates) — use it, adapt it, ship it.
View or download the full playbook on GitHub →
Written by Michael Taus, with Claude (Anthropic). Licensed CC BY 4.0 (text) / MIT (code). If you build on it, a credit and a link back are all I ask — and I’d genuinely love to hear what you change.