How to write prompts that get great answers from Llama

Llama answers from only two things: what it learned in training and what you put in your message. It cannot look anything up, it does not know your situation, and it does not remember your other conversations. So the quality of the answer depends heavily on the quality of the request. The good news is that writing a strong prompt is a skill you can learn in an afternoon. This guide gives you a four-question method, side-by-side examples and the follow-up moves that rescue weak drafts.

Answer four questions before you press send

Think of Llama as a capable new colleague on their first day: quick and well read, but knowing nothing about your project. Before you write, answer these four questions in the prompt itself.

1. What is the job? Start with an action verb: draft, shorten, explain, compare, list, fix, translate, plan. One clear job per message works best.
2. What should it work from? Paste the text, the numbers, the error message or your notes. Without material, Llama fills the gaps with guesses that only sound right.
3. Who is it for? A teacher, a customer, a ten-year-old, your manager, yourself. The reader decides vocabulary, depth and tone.
4. What shape should the result take? A table, a checklist, three options, an email under 120 words, code only with no explanation.

A simple question needs only the first point: "What is the capital of Peru?" is a perfectly good prompt. A message to a client, a study plan or a product page needs all four.

Weak and strong prompts, side by side

Planning a trip

Weak:
Plan a trip to Lisbon.

Strong:
Plan 3 days in Lisbon for two adults in early spring. We like food
markets, viewpoints and walking; we don't enjoy museums. Moderate
budget. Give one table per day with morning, afternoon and evening,
and a short note on getting between places on foot or by tram.

The first version gets you a brochure. The second gets a plan built around your interests, in a layout you can scan on your phone. Llama cannot check today's opening hours or ticket prices, though: it has no internet access and its knowledge stops at a cut-off date. Confirm those details yourself before you go.

Summarising something you paste

Weak:
summarize this [long report pasted below]

Strong:
Summarize the report below for a busy manager who has two minutes.
Give: 1) the main finding in one sentence, 2) three supporting points,
3) any risks or open questions the report mentions. Use only what is
in the text. If something is unclear, say so instead of guessing.

"""
[paste the report here]
"""

Putting the instructions first and the material between triple quotes keeps the two apart. The line "use only what is in the text" makes invented details less likely.

Writing for customers

Weak:
Write a product description for my candle.

Strong:
Write a product description for a handmade soy candle: 200 g, scent
fig and cedar, sold in a reusable ceramic cup. Audience: people buying
a gift online. Tone: warm, not cheesy. Give me 3 versions: about 30
words, about 80 words and about 150 words.

Prompts for images with Llama 4

Both Llama 4 models in our chat, Maverick and Scout, can read a picture you attach: a photo of a whiteboard, a chart, a menu in another language, a screenshot of an error. You can add one image per message. The same rule applies as with text: say exactly what you want done with the image.

Weak:
[photo] what is this

Strong:
[photo of a handwritten shopping list] Type this list out as plain
text, grouped by supermarket section. Mark any word you cannot read
with (?) instead of guessing.

Asking Llama to flag what it cannot read matters, because models sometimes "read" blurry text by guessing. Llama 3.3 70B and Llama 3.1 8B are text-only; if you attach an image while one of them is selected, the chat switches to Llama 4 Maverick. Our guide on choosing the right Llama model explains the differences.

Five patterns worth knowing

Give it a standard to meet

Instead of "act as an expert", say which standard the expert would apply: "Review this cover letter the way a recruiter scanning fifty applications would, and point out anything that would make you stop reading." A role helps only when it changes what the model pays attention to.

Show, don't describe

Words like "professional" or "fun" mean different things to different people. Paste two or three lines you like and write "match the rhythm and vocabulary of these examples". Llama copies a pattern far more reliably than it follows an adjective.

Ask it to interview you first

For bigger tasks, end with: "Before you write anything, ask me up to five questions." You get a better first draft, and you notice what you forgot to mention. Try the interview-me prompt in the chat.

Ask for options, then choose

"Give me six headline ideas in different styles" is faster than rejecting one headline six times. Pick the best one and ask for variations on it.

Ask for steps on hard problems

For maths, logic and debugging, add "work through it step by step, then give the final answer on its own line". You can then check each step instead of trusting a single number at the end.

Steer with follow-ups instead of starting over

The chat remembers the conversation you are in, so the second message is often where the real work happens. Follow-ups that fix most drafts:

• "Cut this by half. Keep the facts, drop the warm-up sentences."
• "Paragraph two is right. Rewrite the others to match its tone."
• "You assumed I have a car. I don't, so redo the plan."
• "List the assumptions you made so I can correct them."
• "Turn this into a table with columns for option, cost and effort."

In our chat you can also edit an earlier message and resend it, or regenerate an answer and switch between versions. Editing is better than piling on corrections when the first prompt was off from the start. When a long conversation starts going in circles or forgetting what you said at the beginning, start fresh: ask Llama to sum up the key decisions so far, then paste that recap into a new chat. Very long conversations are trimmed before they reach the model.

Small adjustments for each model

• Llama 3.1 8B is the smallest model in our chat. Give it one clear task at a time and simple formats. It suits quick rewrites, short answers and brainstorming.
• Llama 3.3 70B is a strong text-only all-rounder. Numbered requirements help it cover every point you asked for.
• Llama 4 Maverick (the default) and Llama 4 Scout also take images. Name what you want from the image in the same message.

Whatever the model, write in the language you want back. According to Meta's model cards, Llama 4 officially supports 12 languages and Llama 3.3 and 3.1 support 8. Other languages often work too, with more mistakes.

Common traps

• Asking about recent events. Llama has no web access. For news, prices, timetables or anything after its cut-off, use a current source.
• Leading questions. "Why is plan A better?" invites agreement. Ask "Compare plan A and plan B, including the weaknesses of each" instead.
• Several unrelated tasks in one message. Split them up, or number them clearly.
• Sharing more than needed. Replace names, addresses and account numbers with placeholders such as [CLIENT]. Our privacy and safety guide explains why.
• Treating confident answers as checked facts. Fluent is not the same as correct; see how to fact-check AI answers.

FAQ

Is there a perfect prompt formula?

No. The four questions (job, material, reader, shape) cover most cases, but the best prompt is the one that gives Llama the details only you know. Clear beats clever.

Does my prompt have to be in English?

Not at all. Llama is instructed to answer in whatever language you write in, so use the one you want back. If you want a different language, say so, for example "Answer in Portuguese".

Why does the same prompt give different answers?

Language models choose each next word with some randomness, so the wording changes from run to run. Regenerate a few times and keep the best version. Our guide to how large language models work explains why.

Can Llama remember my preferences between chats?

No. Each conversation starts fresh. Your history is saved in your browser so you can reopen old chats, but a new chat knows nothing about them. Paste your standing preferences at the start, or pick one of the assistants, which come with instructions already set.

How long can a prompt be?

Long enough to include what matters; a few paragraphs of context is normal. For very long documents, send them in parts and ask for a summary of each part, because the chat limits how much text goes into a single request.