Grok prompts that work: a brief, examples and fixes

A prompt is a brief, the same kind you would give a freelancer you have never met. Grok is quick and widely read, but it knows nothing about your situation, cannot look anything up in our chat and does not remember your other conversations. Everything it needs has to be in the message. This guide gives you a four-part brief, a set of weak prompts rebuilt into strong ones, and the follow-up lines that save an answer that missed.

The four parts of a useful brief

You rarely need all four for a quick question. For anything you plan to send, publish or rely on, check that your message covers them.

1. Goal. What you want done, in one verb if possible: rewrite, compare, plan, explain, check, shorten. Add why if the reason changes the result ("so my manager can decide by Friday").
2. Context. The facts Grok cannot know: your draft, your numbers, the situation, what you already tried. Paste the actual text instead of describing it.
3. Constraints. Length, tone, reading level, things to avoid, a budget, a deadline. Constraints are where generic answers turn specific.
4. Output. The form you want back: a table, five bullet points, a draft email, a numbered plan, code only.

A handy habit is to write the goal first and the output last. The material goes in the middle, set apart with a line of dashes or triple quotes, so Grok can see where your instructions end and your text begins.

Weak prompts, rebuilt

A decision you are stuck on

Weak:
Should I buy an e-bike?

Strong:
I commute 9 km each way, mostly flat, with one steep hill at the end.
I have no car and a small flat with no lift. Compare an e-bike, a
regular bike and a monthly bus pass on cost over two years, effort,
storage and bad-weather days. End with the option you would pick
for me and the one thing that would change your mind.

The weak version produces a generic list of pros and cons. The strong one gives Grok your distance, terrain, housing and alternatives, so the answer is about you. Asking what would change its mind exposes the assumption the recommendation rests on.

A text that needs fixing

Weak:
make this better [paragraph]

Strong:
Edit the paragraph below for a company newsletter. Keep my meaning
and my examples. Cut it to about 80 words, use short sentences and
remove jargon. Then list each change you made in one line.
---
[paragraph]

"Better" can mean shorter, warmer, more formal or simply different. Naming the reader and the length tells Grok which kind of better you want, and the list of changes lets you reject the ones you dislike.

Learning something new

Weak:
Explain inflation.

Strong:
Explain inflation to someone who has never studied economics. Use one
everyday example about groceries, then one about wages. Keep it under
200 words and finish with two questions I could answer to check that
I understood.

You can send the inflation prompt as it is and swap the topic for your own.

Prompts with pictures

Every model in our chat can read an image you attach, one per message. A picture saves a lot of typing, but Grok still needs to know what to do with it.

Weak:
[screenshot] help

Strong:
[screenshot of a spreadsheet] Column C should be the total of A and B,
but some rows show an error. Tell me what is likely wrong, and give
me the formula to put in C2 that I can copy down.

Good image prompts say what the picture shows, what you need from it and what to do with unreadable parts ("mark anything you cannot read with a question mark"). Grok reads and describes images in our chat; it does not draw or edit them.

Patterns that make Grok more useful

Set a word budget

"In 120 words" or "three bullets, no intro" does more for readability than any instruction about style. Grok tends to be thorough; a budget forces it to choose what matters.

Ask for questions before the answer

For a plan, a speech or a CV, end with "Before you start, ask me the questions you need answered." You get a sharper first draft and you notice what you forgot to say. Here is a questions-first starter.

Request several options

Ask for five subject lines in different tones, or three ways to open a talk. Choosing between real options is faster than describing what you want in the abstract.

Add a critic pass

After a draft, ask: "Now read this as a sceptical reader would. What is weak, unclear or unsupported?" Grok is often better at finding problems in a text than at avoiding them on the first try.

Let it say it does not know

Add "If you are not sure, say so" to factual questions. It will not stop every mistake, but it makes doubts visible instead of hidden behind a confident sentence.

Follow-ups that rescue a weak answer

The conversation is the real tool. Instead of starting over, steer:

• "Keep the structure, halve the length."
• "The second option is closest. Give me three more like it."
• "You assumed a team of five. It is just me; redo the plan."
• "What did you assume that I did not tell you?"
• "Rewrite this for someone reading on a phone between meetings."
• "Put this in a table: option, cost, time, risk."

If the first answer missed completely, edit your original message and resend it rather than stacking corrections. In a long chat that has drifted, ask Grok for a five-line summary of what you decided and paste it into a fresh conversation.

Match the prompt to the model

• Grok 4.20, the default, answers right away. It suits drafting, rewriting, explaining and brainstorming. Clear constraints matter most here.
• Grok 4.3 Thinking works the problem through before replying. Use it for multi-step maths, logic puzzles, schedules with many rules and comparisons with trade-offs. Give it all the conditions in one message.
• Grok 4.7 is the newest model in our chat and has a cap of 10 messages a day. Save it for the hardest coding or analysis jobs, and make each of those messages complete.

Not sure which one you need? Our comparison of Grok 4.20 and Grok 4.3 goes through examples.

Mistakes that waste your daily messages

• Asking about today. Our chat has no internet access, so prices, news, scores and timetables may be out of date or invented. Use a current source.
• Loaded questions. "Why is option A the best?" invites agreement. "Compare A and B, including the weak points of each" gets a fair answer.
• Five tasks in one breath. Number them, or send them one at a time.
• Too much personal detail. Swap real names and account numbers for placeholders; see our privacy guide.
• Trusting fluency. A smooth paragraph is not a checked one. Our guide on fact-checking AI answers shows where to look.

FAQ

Do longer prompts always work better?

No. Length helps only when it carries information Grok could not guess. A focused prompt of three sentences often beats a page of vague instructions. Cut anything that would not change the answer.

Should I write prompts in English?

Write in the language you want the answer in. Grok replies in the language of your message, and you can ask for another one explicitly, for example "Answer in Spanish".

Is it worth telling Grok to act as an expert?

Only if you say what the expert would check. "Review this as a landlord deciding whether to rent to me" changes what Grok looks for; "act as an expert" on its own changes little.

Why do I get a different answer to the same prompt?

Each word is chosen with a little randomness, so wording varies between runs. Regenerate and keep the best version. Our guide to how language models work explains why.

Can I save a prompt as a permanent instruction?

Our chat does not store custom instructions between conversations. Paste your standing instructions at the start of a chat, or pick one of the assistants, which already carry a role.