The Second Answer Is Always Better
The first answer is the average of everything the model has seen. It is competent, general, and sounds like everyone. You read it, think "this is fine", and use it — and later you notice it said nothing that could not have been said about any other topic by any other person.
The first answer is never wrong. It is just never yours.
The fix is not a better prompt. The fix is a second pass — but not the kind where you say "make it better" and get the same thing with stronger adjectives. A second pass that makes the model examine what it just wrote, name what is weak, and rewrite with constraints it did not have the first time.
Four prompts.
1. Make it grade its own work
Read what you just wrote. Grade each section on a scale of 1-5 for specificity: does this section say something only this topic could say, or could it appear anywhere?
For every section below 4, name the exact sentence that makes it generic, and rewrite that section so it could not belong to a different article.
The model is surprisingly good at finding its own filler — better than you are, because it is not tired of reading it yet.
The grading forces a standard it did not have on the first pass. Without it, "make it better" is a direction without a ruler, and you get louder where you needed sharper.
2. Ask what it left out
What did you leave out of this answer that I probably need to know?
Not caveats. Not "it depends". Concrete things: a step you skipped because it seemed obvious, an assumption you made that I might not share, a failure mode that changes the advice.
List them. Then pick the two most important and add them where they belong.
The first draft optimises for coherence. It flows, it reads well, it covers the expected ground. What it does not do is interrupt itself to say "wait, there is a thing you should know." That is what this prompt does.
The instruction to exclude caveats is critical. Without it you get a list of "however, results may vary" — which is the model protecting itself, not informing you.
3. Kill the weakest paragraph
Which paragraph in this piece would I lose the least by deleting?
Delete it. Then look at what is left and tell me: is the piece worse, the same, or better without it?
If it is the same or better, find the next weakest and repeat.
This is how you find the real length of a piece: not by setting a word count, but by removing what does not earn its space.
The model almost never says "the piece is worse." It says "better" or "the same", which tells you that paragraph was always dead weight — it just felt necessary because it was there. Two rounds of this typically cut 20-30% and the piece tightens in a way that no amount of "make it more concise" achieves.
4. Rewrite the opening last
Now that the body is final, rewrite the opening paragraph.
The opening must do exactly one thing: make someone who has read the title want to read the first section. It must not summarise the article. It must not provide context. It must not begin with a question.
Write three options. I will pick one.
The opening written before the body is a guess about what the piece will say. The opening written after the body knows.
This is the single highest-leverage edit in any piece of writing, and it is the one almost nobody makes — because by the time the body is done, you are tired, and the opening is already written, and "it's fine" is easier than "let me rethink this."
Why the second pass beats a better first prompt
A better first prompt narrows the space of possible answers. A second pass lets the model see its own answer and find what is missing — and that is information it did not have when writing the first draft.
You cannot front-load hindsight into a prompt. You can only create it by writing once, reading, and writing again. These four prompts make that loop fast, specific, and worth doing every time.
Never use the first answer. It is the one the model writes when it does not know what you will actually keep.
The four prompts are in the carousel on Instagram.
Want the calm version of AI news like this, once a week? Subscribe to the Sharp AI Hub newsletter →