Your Prompt Is Fine. Your Example Is Missing.

When an answer comes back wrong, almost everyone reaches for the same fix: more adjectives. Make it clearer. Punchier. More professional. Less corporate.

None of those words mean anything. Not to the model, and — if you are honest — not to you either, not precisely enough to act on. "Professional" is a vote, not an instruction. The model resolves it the only way it can: by averaging everyone who has ever used that word. That average is exactly the output you keep rejecting.

The fix is not a better description. It is an example.

Four steps.

1. Show one output you would have accepted

Before you write anything, here is one output I would have accepted: [paste it].

Do not copy it. Tell me what it is doing — structure, sentence length, what it opens with, what it refuses to do.

Then write mine to the same spec.

The instruction that matters is do not copy it. Without that, you get a paraphrase of your example wearing your topic as a costume.

Making it describe the example first is not ceremony. It is a test you can read: if the description is wrong, you have caught the misunderstanding before it cost you a draft. And more often than you would expect, the description tells you something about your own taste you had never put into words.

2. Show one you would have rejected

Here is one I would have rejected: [paste it].

Name the difference between this and the accepted one in a single sentence. Not "it's less engaging" — the actual mechanical difference.

That sentence is the rule. Repeat it back to me before you write.

One example gives a direction. Two give a boundary, and a boundary is what you were trying to describe all along.

The near-miss is worth more than the disaster here. An obviously terrible example teaches nothing — the gap is too wide to be informative. Pick the one that was almost right, the one you rejected on a second read. The line between almost and yes is the entire lesson.

Ban "engaging", "flat", "generic" in the answer. Those are verdicts. You want the cause.

3. Make it state the rule

From those two examples, write the rule as an instruction someone else could follow without seeing either one.

If your rule needs the examples to make sense, it is not a rule yet — it is a description. Try again.

This is the step that turns one good answer into a repeatable one.

A rule that survives without its examples is portable: it goes in your saved prompt, your project instructions, your team's style guide. A rule that collapses without them ("write it more like the first one") has to be re-taught every session, and you will re-paste those examples for the rest of your life.

The test is strict on purpose. Most first attempts fail it, and the second attempt is usually the one worth keeping.

4. Test it on something new

Now apply that rule to this, which neither example covers: [new input].

Then tell me where the rule stopped being enough, and what you did instead. That gap is the next example I owe you.

A rule that only works on the cases it was derived from is not a rule — it is a memory.

The last question is the compounding one. Every time you run this, the model tells you exactly which example is missing from your set. Add it. Three or four rounds in, you stop writing prompts and start maintaining a small library of decisions you have already made.

Why this works when adjectives do not

An instruction describes the target. An example is the target.

Everything you actually care about — rhythm, how much it explains before it commits, whether it hedges, where it stops — lives in the space between sentences, and there is no vocabulary for most of it. You have been trying to compress a taste into a word, and taste does not compress.

So stop paying the compression cost. Show the thing.

The whole method is one sentence: instructions describe, examples decide. If you find yourself typing "make it better", you already know what to paste instead.

---

The four prompts are in the carousel on [Instagram](https://instagram.com/sharpaihub).

Want the calm version of AI news like this, once a week? Subscribe to the Sharp AI Hub newsletter →