Why You're Burning Tokens in Three Turns

“You’ve exceeded your token usage — buy more to continue.”

If you build anything with AI tools (ChatGPT, Claude, Gemini), you’ve seen some version of that line. Four turns into a session, the meter climbs, and then the wall. The session craters, you start again — same work, fresh wallet — and by the third reload you’re swearing at the model, the framework, the rent.

The advice underneath those posts is always the same: smaller model, chunk your prompts, switch providers. None of it touches what’s actually broken.

I wrote a short piece this week about the real problem. The operators bleeding tokens have been taught the wrong physics. They’re standing where a customer stands — asking to be served — when what they’re actually doing is trying to build. The model treats them like customers because that’s where they’re standing. The mismatch eats the budget.

It’s the same thing running a kitchen has always required: lay the substrate down before service starts, so the session reads what you wrote instead of re-deriving it every time. Short read. It names the thing nobody on those threads is naming.

Read it here → Why You’re Burning Tokens in Three Turns

Tony Cooper tonycooper.co.uk

tony.cooper@webuildstores.co.uk

P.S. If you’re running sessions that collapse in three turns and you can’t see why, reply and tell me what you’re building. The gap is almost never the model.