The failure mode with code generation is not bad syntax. It is a confidently written 200-line implementation of the wrong approach.
The fix is two prompts:
Prompt one: "Here is the problem and the constraints. Do not write code yet. Give me two or three approaches with the trade-off of each, and say which you would pick."
Read it. Disagree if you disagree. Then:
Prompt two: "Go with approach 2. Now write it."
This costs you thirty seconds and saves the review cycle where you realise on line 180 that the whole thing assumes single-tenant.
It also surfaces assumptions early. Half the time the plan reveals the model misread a requirement, and that is much cheaper to fix in a plan than in code.
Attached to the prompt(2)
10.2k2
Comments (2)
No comments yet. Be the first to share your thoughts!