An 8-person startup writes step-by-step guides for its AI agents

AMP, an AI company run by cofounder and CEO Rachel Woods, has four full-time staff and four part-time contractors. It says it would need four to five more people without AI. The company does not let its AI agents work on guesswork. Each agent follows a written step-by-step guide, which the company calls a playbook.

What a playbook is

Every department at AMP has its own system. Each system is a group of playbooks, and the AI agents in them can run alone or together. Woods says to start with one task you do over and over, then write it out the way you would train a new hire.

• Write what starts the task, what goes in, and what should come out.
• List each step, the decisions in it, and what good looks like.
• Run it with AI and compare the result to how you would have done it.
• Fix the playbook, not the output, until you trust it to run without you.

What it does at AMP

The sales playbook reviews customer calls to check that they meet company standards. A sales manager would otherwise do that by hand. Sales staff also ask the playbook their questions instead of asking a manager. It pulls answers from the team's own guide. The roles AMP says it would otherwise hire for are a sales and marketing manager, a customer support specialist, an executive assistant and an operations coordinator.

What it costs

AMP pays for two kinds of AI. Each person has a subscription to Claude and ChatGPT. The playbooks also run in the background without anyone starting them. In September, AMP spent $8,510 on Claude Code for that work. It is billed by usage and is separate from the subscriptions.

Why the tool matters less

Woods says the company follows the idea of owning the playbook and renting the technology. If the playbook is good, it does not matter much which tool runs it. AMP uses Claude Code now. She says the same playbooks could run in ChatGPT, Grok, Gemini or Microsoft Copilot.

She also says people should write the logic themselves. Including steps you do not understand makes the system hard to trust. A person still has to decide and design how the work gets done, she says. AI speeds up the rest.

Who owns what

A few of the team's more advanced AI users own the larger, team-wide playbooks that the business depends on. Everyone also writes playbooks for their own personal workflow. Those personal playbooks serve as the testing ground for what gets promoted to the wider team.