Describe the Gap

Say What Is Missing, Not That It Is Wrong

The single most productive follow-up habit is describing the gap between what you got and what you wanted, precisely, in one message. "That is not right" contains almost no information. "You covered the process but I needed the decision — who signs this off and when" contains all of it. "Too long" is weak; "cut the second and third sections entirely and halve the rest" is actionable. "Wrong tone" is weak; "it reads like marketing — I need it to read like an internal note between two engineers" is actionable. You are not complaining, you are specifying. Each round should reduce the remaining distance, and it only does that if you named the distance.

• Weak: "That's not what I wanted." Strong: "You gave me the process; I need the decision and the owner."
• Weak: "Too long." Strong: "Cut sections two and three; halve the rest."
• Weak: "Wrong tone." Strong: "Reads like marketing; I want an internal engineering note."
• Name what to keep as well as what to change, or the good parts get rewritten too

Iterate, Do Not Restart

Starting over with a fresh prompt after every imperfect answer throws away the useful part. The output you just got is information: it shows you what was ambiguous, what defaults got filled in, and which of your assumptions were never stated. Keep it and correct it. Say what to preserve — "keep the structure and the first section exactly as they are" — before you say what to change, otherwise you will get a wholesale rewrite that loses the bit you liked. There is a limit, though. If a thread has accumulated five corrections and each answer is now worse, the transcript itself has become the problem, and the right move is a clean chat with one good prompt containing everything you learned.

• The imperfect answer tells you what your prompt failed to specify — use it
• Say what to keep before what to change, or the good parts get lost
• Correcting is usually faster than rebuilding the prompt from scratch
• After about five corrections, restart clean with everything you have learned

Louder Does Not Work

When something is not landing, there is a strong instinct to escalate: repeat the same prompt, put the instruction in capitals, add "IMPORTANT" three times, tell it this really matters. Mild emphasis can help a genuinely buried instruction get noticed, but it does nothing about the actual problem, which is almost always that the instruction is unclear, contradicted by something else in your prompt, or asking for something the model cannot reliably do. Repetition without change is the same experiment run twice. If an instruction is being ignored, look for the contradiction first — very often another line in your own prompt is quietly pulling the opposite way, and no amount of shouting resolves a conflict.

• Repeating an unchanged prompt is the same experiment, not a fix
• Capitals and "IMPORTANT" do not resolve an unclear or contradicted instruction
• Check your own prompt for a line pulling the opposite way — this is the usual cause
• If an instruction is ignored three times, it is probably ambiguous, conflicting, or impossible

Try It Yourself

The next disappointing answer you get is the exercise. Do not retry it louder — describe the gap once and watch what changes.

That is not quite what I needed.

What you gave me: [one line]
What I actually need: [one line]
The difference: [name the specific gap]

Revise only that part — keep everything that already works.