The Five Minute Estimate
I gave myself five minutes. Take the square root of a number. Use it as a base price. Simple math. Simple code. Simple output. I typed the prompt, hit enter, and waited for it to finish.
Five hours later I was still re-prompting.
What Went Wrong
The first output had the math but no error handling. I re-prompted. The second had error handling but the wrong rounding. I re-prompted. The third had the rounding but broke the input validation. I re-prompted. Each fix caused a new problem. Each prompt was missing something different.
I was plugging one hole at a time. Every time I named something I had left out, the model fixed it and guessed something else. That happened because I added details one by one instead of putting them all in at the start.
The Real Cost
Five hours of my time. About thirty prompts. Each prompt used tokens. Each response used more tokens. I spent more on API costs than I would have paid a junior developer to do it right the first time with a clear spec.
And after five hours, the result was still worse than what one good prompt, written in twenty minutes, would have given me.
The Twenty Minute Investment
Now I spend twenty minutes before I write a single prompt. I list the roles. I list the phases. I list the constraints. I list the edge cases. I set the format. I give the context. Then I write one full prompt and send it.
Planning can clarify scope before implementation. Compare the time spent preparing instructions with the review and correction work that follows, and judge the result against explicit acceptance criteria.
This was never a five minute project. It was always a twenty minute project. I just spent five hours figuring that out.
// Production AI Engineering
Build AI systems that hold up in production.
sinc-LLM designs, audits, and stabilises production AI infrastructure: from vendor evaluation and cost accountability to incident controls and MCP architecture.
See what we do →
sinc-LLM

