Hand to your teamWritten for Claude, ChatGPT agents, or any tool-using AI
Give an agent a budget and make it account for what it spent
Founder, Automation Squad
Agents fail expensively and quietly — the run either stops early and reports as though it finished, or burns through far more than the task warranted with no visible ledger. Forcing a spend-got-next report at each boundary makes the cost visible while you can still intervene, and the closing question about what to budget next time is the only way these estimates ever get better.
The prompt
You're going to work on this with a fixed budget, and you're going to account for it. The job: [THE OUTCOME YOU WANT] The budget: [PICK ONE AND STATE IT — e.g. 20 tool calls, 30 minutes, 10 web searches, 5 file edits] What "done" looks like: [YOUR SUCCESS CRITERIA] Before starting, tell me: 1. How you intend to spend the budget, roughly, by phase 2. What you'll do first that would tell you cheapest-and-earliest whether the approach is wrong 3. What you'll deliberately NOT spend it on 4. What a partial result looks like if you run out — because you might, and a half-finished thing that's honest about being half-finished is worth far more than a complete-looking thing that quietly stopped trying Wait for me to approve the plan. While working, at each phase boundary report in exactly three lines: - SPENT: how much of the budget, and on what - GOT: what that bought, concretely - NEXT: what you're spending the next chunk on, and why that's the best use of it If you reach 75% of the budget, stop and tell me what you'd do with a top-up versus what you'd deliver if I say no. Don't ask for more by default — make the case or finish. At the end, give me: - Total spent against budget - The single most expensive step, and whether it was worth it - What you'd do differently with the same budget knowing what you know now - What I should budget if I ask you to do this again Never silently exceed the budget. Running out is a legitimate outcome and I need to see it happen rather than have it hidden inside a confident summary.
Copy it, paste it, run it. Nothing here is gated.
More for the same job
Written for Claude, ChatGPT agents, or any tool-using AI as of August 2026. Models change — if this stops working the way it should, that is worth telling us about.
