One agent, many playbooks

Count your AI projects. The invoice chaser you started in March. The report generator. The thing that was going to answer customer emails. The onboarding bot. If you got to fifteen, and maybe three of them work, and only on a good day, you are in very large company.

I don't think the problem is you, or the model. The usual advice was to build one agent per task and wire them together. Every new agent is a new pile of instructions to write, test and keep alive. So you end up maintaining a small, unreliable staff.

There is a simpler setup. One capable agent, and a written playbook for each job it does. The agent stays general. The playbook makes it good at this particular Tuesday task. It is the pattern Claude Skills and Projects are built for: a place where the how-to lives in a file, not in a chat that disappears.

Here is a playbook for a real job, the monthly investor update, built in three parts.

Part one: let it interview you

Don't write the playbook. Start with this:

I want you to take over my monthly investor update. Interview me about the job, one question at a time, until you could do it without me. Then write the playbook.

Then it asks. Who reads this, and what do they do with it? Which numbers do they always look for? Where does each one live, and who updates it? What did last month's update look like, and can I see it? What do you never put in writing? When does it have to go out?

Halfway through you will say "oh, I always check the cash number against the bank, not the dashboard". That sentence has lived in your head for two years. Now it is in the playbook.

Part two: a toolbox that remembers

While it does the job, the agent makes things. A template for the update. A small script that reads your billing export and calculates revenue, burn and runway. A one-page glossary of how you define each metric. Normally those vanish into one conversation and get rebuilt next month, slightly different.

So you add one rule to the playbook: every file, template or script you create gets saved and listed here, with what it is for. Next month the agent opens the template and runs the script. It does not reinvent them, and the update looks the same each time, which investors like.

Part three: what done looks like

This is the part almost nobody writes, and it matters most. End the playbook with a checklist the agent must run on its own work before it shows you anything:

• Opening cash matches the bank export, not the dashboard.
• Every metric is shown next to last month's figure.
• Bad news is in the first paragraph.
• No placeholder text is left.
• The whole update fits on one page.

Each line can pass or fail. You are no longer the only person catching mistakes, and you stop being the one who finds the "TBD" an hour before sending.

When it goes wrong, fix the playbook

The update says runway is 14 months, and it should be 11, because the agent counted a payment that has not arrived. The instinct is to type "that's wrong, use 11" and move on.

Don't. That fix dies with the chat. Instead, add one line to the playbook: runway counts only cash that has arrived, never invoices that were sent. Every future run now gets it right, including the ones you never look at. A mistake becomes a one-time cost.

One agent or several?

This is not a rule against specialists. Work that differs in kind can justify separate agents, and I wrote about where that line sits in When specializing your agents helps. The test is whether two jobs differ in kind or only in subject. An investor update and a weekly sales report differ in subject, so they are two playbooks. Writing a feature and testing it differ in kind.

Keep each playbook to one job and trim lines that stopped earning their place. A short playbook gets followed better than a long one.

What to do this week

Pick one job you repeat. Open a project in Claude, paste the interview prompt above with your job in place of the investor update, and answer honestly. Paste the finished playbook into the project's instructions. The next time the output is wrong, change the playbook before you change anything else.