I'm not a programmer. This fall I started turning Claude Code into an assistant that actually runs parts of my day: it plans, it remembers what we decided, it drafts the emails, and it asks me before it does anything that can't be undone.
Here's the setup that made it work, in plain words.
A chat forgets. An assistant can't. The fix is simple: a folder of notes the assistant reads and writes. One note for each ongoing job, a daily note for what happened today, and an index that says where everything is.
The important part is that the assistant loads only what the current job needs. It doesn't read everything every time. It knows where things are and pulls them when it needs them.
Don't build a new bot for every task. Give one assistant a playbook per job. Each playbook has three parts:
That last one matters more than it sounds. This week my assistant's own checklist caught a bug in our email tool that would have unsubscribed the exact customers who wrote back to us.
An assistant that can send, post or spend money needs a lock. Mine asks my phone first: one plain sentence about what it wants to do, and a yes or no button. One yes covers one action. It can't change the lock or go around it.
Most of my work with it happens from my phone: a voice note while I'm out, and it does the job back at the computer and answers. That's when it starts feeling like a chief of staff instead of a chatbot.
Want the starting point? The free Chief of Staff starter kit sets up the memory, the daily plan and the first playbook. Add the phone line, the approval lock, or the email line when you want them.