The one-sentence method
Don’t just tell AI to write code. Clarify requirements first, then propose a solution, write a plan, execute in stages, and finally prove the result correct with tests and acceptance criteria.
Why not just tell AI to write
Handing a task to AI in one sentence produces code that “looks right”: no requirement clarification, so boundaries are guessed; no proposal review, so architecture is arbitrary; no acceptance criteria, so correctness is unprovable. Processed input buys verifiable output.
The five-step SOP
- Clarify: pin down vague requirements, define boundaries and success criteria
- Propose: have AI present 2-3 approaches with trade-offs, then write the design
- Plan: decompose the design into independently verifiable tasks, each with tests
- Execute in stages: one task at a time, verifying each step
- Accept: prove correctness with tests and acceptance criteria — not “looks fine”
The experiment framework
The value of this SOP can be validated with a controlled experiment: the same task, one group “hands it to AI in one sentence”, the other “feeds it through the SOP” — compare completion time, rework count and acceptance rate. The goal is a process that can be promoted within a team, backed by data.
Lessons from practice
- Time spent clarifying is repaid tenfold during execution
- If you cannot write acceptance criteria, you have not thought the requirements through
- An SOP is not a constraint — it turns luck into predictability