How to Write Better ChatGPT Prompts: A Before-and-After Checklist

To write better ChatGPT prompts, say exactly what you want done, give the context the model cannot know (who it is for, why, what constraints apply), specify the format of the answer, and paste in the real material instead of describing it. Then iterate: treat the first answer as a draft and tell the model precisely what to change.

Everything below works the same way in Claude, Gemini, Copilot and other assistants. The checklist is ordered by how much each item tends to improve results, with a before and after for each.

#1. Did you say what the output is for?

The single biggest upgrade. Purpose tells the model what to optimize.

Before:

Summarize this article.

After:

Summarize this article for my manager, who will decide whether our team
should attend this conference. She cares about cost, which sessions
cover data engineering, and whether talks are recorded. Five bullet
points max.

[article text]

The first prompt gets a neutral summary of everything. The second gets a decision memo, because now the model knows what a useful summary would contain.

#2. Did you include the material, or just describe it?

A lot of weak answers come from asking the model to work on something it cannot see.

Before:

Make my cover letter sound more professional.

After:

Here is my cover letter for a junior data analyst role at a mid-size
logistics company. Make it sound more professional without making it
stiff. Keep my examples about the warehouse dashboard project, they
are the strongest part.

[cover letter]

Obvious when written down, but people ask for help with "my code", "the email", "this situation" constantly. Paste it in. Also mark which parts must survive the edit, or the model may "improve" your best paragraph away.

#3. Did you name the audience?

"Explain recursion" gets a very different answer when you add "to a 12-year-old who knows basic Scratch" vs "to a developer who has used loops for years but never needed recursion". Audience sets vocabulary, depth and which examples are worth using.

#4. Did you specify the format?

If you will paste the answer somewhere, say what shape it needs.

Before:

Give me ideas for a team offsite.

After:

Give me 8 ideas for a one-day team offsite for 12 software engineers in
Chicago in October. Budget about $100 per person. Mix of active and
low-key options.

Format as a table with columns: idea, rough cost per person,
why it works for this group. No intro paragraph.

"No intro paragraph" is worth adding more often than you would think. Assistants love to open with a sentence restating your question.

#5. Did you set constraints the model would otherwise guess?

Length, tone, reading level, what to avoid, what must be included. Each one prevents a round of corrections.

Useful constraints are specific. "Keep it short" is vague; "under 80 words" is not. "Make it engaging" means nothing to a model; "open with the problem the reader has, not with our company name" does.

#6. Did you show an example when the style is hard to describe?

If you want something in a particular voice or format, one or two examples beat a paragraph of adjectives.

Write 5 more product taglines in the same style as these:

- "Coffee for people with deadlines."
- "Shoes that don't need a speech."

Products: a note-taking app, a bike lock, a meal-prep container,
a standing desk, a password manager.

Vary your examples, though. If both examples are four words, you will get four-word taglines. More on that in zero-shot vs few-shot prompting.

#7. Did you ask it to think before answering a hard problem?

For multi-step questions (planning, math, comparing options against several criteria), ask for the reasoning first and the conclusion last.

Before:

Which of these three apartments should I rent?
[details]

After:

Help me choose between these three apartments. My priorities, in order:
commute under 30 minutes to downtown, rent under $2,200, in-unit laundry.

For each apartment, check it against each priority. Then recommend one
and explain the trade-off I'd be making.

[details]

You gave it the criteria and the order of work, which matters more than the generic "think step by step". Newer reasoning modes do some of this automatically. See chain-of-thought prompting explained for when it helps.

#8. Did you let it say "I don't know"?

Models fill gaps with plausible guesses unless told they can stop. Add a line like:

If you're not sure about a fact, say so instead of guessing.
If the document doesn't mention something, say that it doesn't.

This does not make the model perfectly honest, but it gives it an acceptable alternative to inventing an answer. For factual work, also ask for sources and then check them. Citations from chat assistants can be wrong, especially when they are not browsing or reading documents you provided. Grounding answers in real documents is a topic of its own; see the grounding topic.

#9. Are you iterating, or starting over?

When the first answer misses, the instinct is to rewrite the whole prompt. Usually a targeted follow-up works better:

Good structure. Three changes:
1. The second paragraph is too salesy, make it factual.
2. Cut the last bullet, it repeats the first.
3. Replace "leverage" with plain words throughout.

Specific, numbered feedback is easy for the model to apply. "Make it better" is not.

The exception: if a conversation has gone in circles for many turns, start a fresh chat with a consolidated prompt that includes everything you have learned. Long, messy conversations accumulate contradictory instructions.

#10. Did you ask it to ask you questions?

For tasks where you are not sure what context matters, flip it:

I want to write a project proposal to get budget for a new analytics
tool. Before writing anything, ask me the questions you need answered
to make this proposal convincing. Ask them all at once.

This is one of the most underused moves. The model's questions often surface context you did not think to give, such as who approves the budget and what they objected to last time.

#A combined example

Putting most of the checklist together on one realistic request:

Before:

Write a LinkedIn post about our product launch.

After:

Write a LinkedIn post announcing that our invoicing app now supports
recurring invoices.

Audience: freelancers and small agency owners who already follow us.
Goal: get existing free users to try the feature this week.
Facts: available on all plans, set up in Settings > Invoices > Recurring,
supports weekly, monthly and custom schedules.

Tone: plain and friendly, no hype words like "revolutionary".
Length: under 120 words. One short line at the end inviting questions.
No hashtags.

Here is a previous post of ours whose tone I liked:
[post]

Nothing in the "after" version is a trick. It is the information a good human copywriter would ask for before starting.

#What doesn't matter as much as people think?

Magic phrases. Lines like "you will be tipped $200" or "take a deep breath" circulated widely. Some showed effects on specific older models in specific tests; none replace a clear task and real context.

Elaborate role prompts. "You are a world-renowned expert with 30 years of experience" adds less than one sentence about the actual audience. Roles help most when they bring information, for example "You are reviewing this as a security engineer, focus on authentication".

Capital letters and threats. Current models follow instructions closely. Shouting tends to make them overapply a rule rather than follow it more accurately.

#Does this work the same in Claude and Gemini?

The fundamentals transfer completely. The details differ a bit, and the companion piece on prompting ChatGPT vs Claude vs Gemini covers what to adjust. If you want to see where each habit fits in the bigger picture, what is prompt engineering is the overview, and the fundamentals topic turns these checks into practice questions.