The difference between a bad and good prompt (5 before/after examples)

The fastest way to get better at prompting is not to learn ten frameworks. It is to look at five prompts you have actually written, see what is wrong with them, and watch what changes when you fix it.

What follows is exactly that. Five common situations, written first the way most people write them, then rewritten with three small additions: context, constraints, and a concrete example or two. After each pair, a short note on what is doing the work.

1. Drafting an email

Weak: Write me an email to my landlord about the heating not working.

You will get something — but it will be generic, in someone else's voice, longer than you need, and not yet useful.

Better: Write a polite but firm email to my landlord. The central heating in my apartment has been broken for two weeks. I have messaged him twice on WhatsApp without a real response. I would like the issue fixed within five working days. My tone is normally direct but not aggressive — I am not threatening anything yet, just being clearer.

>

Constraints: under 120 words, three short paragraphs, no exclamation points, no apologies, no "I hope you're well."

>

End with a clear next step.

What changed: who is sending it (you, your tone), what has already happened (two unanswered messages), what outcome you want (five working days), and a length cap. The model now produces a usable first draft instead of a placeholder.

2. Summarizing a long document

Weak: Summarize this 30-page contract.

You will get a polite, hedged, generic summary that often misses what you actually care about.

Better: I am about to sign this 30-page contract as a freelance designer in Estonia. Read it carefully and tell me:

>

1. The three clauses that most affect my obligations. 2. Anything that gives the other side a unilateral right (terminate, change pricing, claim IP). 3. Anything that contradicts itself or sounds unusual for a contract like this. 4. The two questions I should ask my lawyer before signing.

>

Mark anything you are unsure about with [unclear].

What changed: you told the model who you are, what you care about, what shape the answer should take, and gave it permission to flag uncertainty. The result is closer to "advice from a careful friend" than "a generic summary."

The [unclear] tag matters more than it looks. Without it, the model fills gaps with plausible-sounding guesses. With it, you get back a summary that admits its own limits.

3. Generating ideas

Weak: Give me ideas for a side project.

You will get a list of generic side projects that could be for anyone. None of them will feel like yours.

Better: Help me brainstorm a side project. About me: I am a product manager at a B2B SaaS company in Tallinn, I write a lot, I like teaching, and I have about 5 hours a week. I am not trying to make money fast — I want something I would still enjoy in twelve months. I have already considered and rejected a newsletter (too lonely) and a podcast (don't have the equipment).

>

Generate ten ideas. For each, give me a one-sentence pitch, the kind of person who would care, and the single hardest part of doing it. Sort them from "easiest to start this weekend" to "most ambitious."

What changed: who you are, how much time you have, what you have already ruled out, and the format of the output. The model now has constraints to push against, and the ideas it returns will feel pickable instead of generic.

The phrase "I have already considered and rejected X" is criminally underused. It stops the model from suggesting the obvious things and forces it to actually think.

4. Asking for an explanation

Weak: Explain machine learning.

You will get a textbook paragraph that you have read a hundred times.

Better: Explain machine learning to me three different ways, in roughly 200 words each:

>

1. To my 12-year-old nephew who is curious but easily bored. 2. To a senior business stakeholder who keeps asking "but how is this different from a spreadsheet." 3. To me — someone who has used ChatGPT a lot but never studied any of the underlying ideas — to give me a working mental model.

>

Use one short concrete example in each. Do not include the words "algorithm" or "artificial intelligence" in version 1.

What changed: you asked for three calibrated explanations, gave the model audiences to write for, and even removed permitted vocabulary in one of them. By the third version, something will click that did not before. This pattern — "explain it three ways" — works for almost any concept you are partly stuck on.

5. Getting unstuck on a decision

Weak: Should I take the new job?

You will get a balanced, hedged "it depends" answer with a list of generic things to consider. Useless.

Better: I am deciding whether to take a new job. Before you say anything, ask me five questions you would need answered to give me a useful opinion. Wait for my answers before continuing.

>

Once I have answered:

>

1. List the three strongest arguments for taking it and the three strongest arguments against. 2. Tell me what evidence would flip your view. 3. Then, only after all of that, tell me what you would actually recommend, and how confident you are.

>

Push back if anything I say sounds inconsistent. Do not give me a polite both-sides answer — I have plenty of those already.

What changed: you forced the model to interview you before answering, you scaffolded the analysis, and you explicitly asked it to stop hedging. The first reply will be five questions. After you answer, what comes back will be far more like advice from a thoughtful colleague than a generic pros-and-cons list.

The pattern across all five

Look at the five "better" prompts side by side. They share four habits:

1. Context. Who you are, what you have already tried, what you are about to do. The model is much better when it knows your situation.
2. Constraints. Length limits, tone restrictions, formats, forbidden words. Constraints lift quality more than length does.
3. Examples or audiences. "For my 12-year-old nephew," "as a freelance designer in Estonia," "as if I were the skeptical senior stakeholder." Concrete anchors beat vague adjectives.
4. Permission for uncertainty. The [unclear] tag, the "wait for my answers," the "tell me how confident you are." Tells the model you would rather have a calibrated answer than a confident wrong one.

Notice what the better prompts are not: they are not longer for the sake of it, they do not use jargon ("act as a senior level expert..."), they do not stack adjectives. They are clearer about what you actually want.

A useful habit

After any meh ChatGPT answer, pause before re-prompting. Ask yourself: which of those four ingredients did I forget? Add one and try again. Within a week of doing this, you will catch yourself adding them automatically, and you will not need to "learn prompt engineering" the way it is often taught — because you will already be doing it.