Before you hit send
1. Did you say what you want, specifically? "Help with this email" is a request. "Shorten this email to 3 sentences and make the tone more direct" is an instruction. AI responds much better to the second kind.
2. Did you give context it can't guess? Who's the audience? What's already been tried? What constraint matters (length, deadline, budget)? If a fact matters to the answer and the AI has no way to know it, say it.
3. Did you specify the format you want back? A table, a numbered list, plain prose, a specific word count — naming the format upfront saves a follow-up round-trip almost every time.
4. Is there a role that would help? "Review this as a skeptical editor" or "explain this like I'm new to the topic" gives the AI a consistent lens to respond through, which often improves relevance more than people expect.
5. Did you ask for exactly one thing? Prompts that bundle three unrelated requests together tend to get a shallow pass at all three instead of a good answer to one. If you have multiple asks, consider separate messages.
After you get a response
6. Is this a draft, not a final answer? Your first prompt rarely needs to be perfect. "Make it shorter," "less formal," or "add a specific example here" are completely normal, expected follow-ups — treat the exchange as iterative.
7. Did you check anything that needs to be true? If the response includes a fact, statistic, citation, or claim that matters, verify it independently before relying on it. See AI Hallucinations Explained for why this step matters.
A quick before/after
Before: "Write a product description."
After: "Write a 60-word product description for a reusable water bottle, for an online store's product page. Tone: simple and direct, no exaggerated claims. End with one practical detail, not a slogan."
Same request, dramatically more usable output — because it answers questions the AI would otherwise have to guess at.
Keep a few templates handy
Our Prompt Library has original, ready-to-use templates already built around this checklist, with placeholders you fill in — a faster starting point than writing every prompt from a blank page.
The same request, written three ways
Abstract advice about specificity is easy to nod along to and hard to apply, so here is one request at three levels of detail.
Weak: "Write about our new feature."
The model has no idea who is reading, how long it should be, what the feature does, or why anyone should care. It will produce something grammatical and useless.
Better: "Write a short announcement about our new bulk export feature for our customer newsletter."
Now it knows the audience and the format. The output will be usable but generic, because it still does not know what problem the feature solves.
Strong: "Write a 120-word announcement for our customer newsletter about a new bulk export feature. Our customers previously had to download reports one at a time, which took them an hour each month. Lead with the time saved, not the feature name. Plain and direct — we do not use exclamation marks."
The third version takes twenty seconds longer to write and removes three rounds of revision. That trade is the entire skill.
When the answer is still wrong
Even a well-built prompt sometimes misses. The instinct is to rewrite from scratch, and it is usually the wrong move — you throw away the parts that worked. Try this order instead:
- Name the specific problem. "The second paragraph is too formal" beats "make it better." The model cannot see what you dislike unless you point at it.
- Ask what it assumed. "What did you assume about my audience?" often reveals exactly where the misunderstanding started.
- Give a counter-example. Show a sentence in the style you want. One example carries more information than a paragraph of description.
- Only then start over, and when you do, carry forward what you learned from the failed attempt.
The checklist in one place
Before you send a request, check that you have supplied: who the reader is, what format and rough length you want, any facts the model could not know, the tone in concrete terms, and what you specifically do not want. Missing one of those is fine. Missing four is why the answer disappointed you.
If you want the underlying structure rather than a checklist, Prompt Engineering for Real Work covers it in four short lessons, and the Prompt Library has ready-made templates already built this way.