How not to go insane (using AI for work)

When you run several AI sessions across different projects, and you're under pressure to be efficient, three frustrations show up.

1. Losing context

You know the AI knows all the important details. You just don't know what they are anymore. They're written down somewhere, in a hundred files and a few thousand lines, in something called "Deliverables-v2-final.md".

Try to end every session with the agent writing a one-page handoff covering what changed, what's decided and what's still open. It is a litmus test of your context awerness - if you don't know what it's talking about, your agent has a different context than you.

2. Losing control

You know what you want to achieve and what you want to avoid. You know the AI knows it too, because you told it.

But do you know every decision, action and assumption it made along the way? And should you?

I don't think you should. Checking every step defeats the point of delegating. Instead, write down the few things that are yours to decide: anything irreversible, anything that touches money or other people or creates commitments, anything that changes the goal. Before any of those, the agent has to stop and ask.

3. Losing satisfaction

Your agent has just completed a complex project flawlessly.

You can tell it it did a good job, but can you also feel proud of yourself?

Pride comes from the struggle: the dead ends, the fix ideas. When the agent does the struggling, it gets the satisfaction.

Keep the architecture piece of every project for yourself. It's ok to ask for feedback but you implement it in your design.

—

None of this is finished. We're all figuring out this kind of work as we go. What's working for you?