Prompting basics

A prompt is the text you give an assistant. Most disappointing answers come from a missing piece of information, not from a weak model. This lesson gives you a simple structure for prompts that work, and a habit of improving them step by step.

Watch

Prompting 101Anthropic · 25 min

Why prompts matter

The model sees only what is in the conversation. It cannot see your job, your audience, your company or what you meant. A short vague request forces it to guess, and its guess will be generic. A prompt that supplies the missing pieces lets it be specific.

Think of briefing a capable new colleague who knows nothing about your situation. What would they need to know to do it right the first time?

The five parts of a good prompt

1. Task: what you want done, as a clear instruction.
2. Context: the background that is not obvious, such as who the readers are and what has already happened.
3. Input: the actual material to work on, such as the text, data or error message.
4. Constraints: what to include, avoid, or keep within a limit.
5. Format: what the answer should look like, such as bullets, a table, JSON or a short paragraph.

Not every prompt needs all five. Check each one and add what the model would otherwise have to guess.

A weak prompt and a strong one

Weak:

Write about our outage.

Strong:

Task: Write a status update about last night's outage.
Context: We run a booking platform used by small clinics. The outage lasted
40 minutes (21:10 to 21:50 UTC) because a database failover failed. No data
was lost. Readers are clinic managers, not engineers.
Constraints: Apologize once, no technical jargon, no blame, under 120 words.
Format: A subject line, then one short paragraph, then one line on what we
are changing.

The second prompt takes a minute longer to write and saves several rounds of fixing.

Be clear and direct

• Say what you want, not only what you do not want. "Write in plain language" works better than "do not be technical".
• Be specific about size and style. "About 100 words" beats "short". "Friendly but professional" beats "nice".
• Give the reason when it helps. "Keep it brief because it goes in a text message" lets the model make good choices you did not list.
• Ask for one thing at a time when quality matters. Do the next step in a follow-up.

Name the audience and the role

Telling the model who the answer is for improves it more than almost anything else: "explain to a finance manager with no IT background" gives a very different answer from "explain to a network engineer".

You can also set a role: "Act as a careful technical editor." A role sets tone and focus. It does not give the model knowledge it does not have, and it does not make false answers true.

Prompting is iterative

The first answer is a draft. Treat the conversation as a loop:

1. Read the answer critically.
2. Say precisely what to change: "shorter, drop the second paragraph, keep the apology".
3. Ask for alternatives when you are unsure: "give me three versions with different tones".
4. When it goes off track, start a new chat with a better first prompt instead of arguing in a long thread.

A quick checklist

Before sending an important prompt, ask:

• Would a new colleague understand what is wanted, for whom, and why?
• Is the input included, rather than described?
• Did I say what a good answer looks like?
• Did I say what to avoid?

Resources

Check yourself

1. Rewrite a vague prompt you used recently so it has all five parts. Compare the answers.
2. Why does naming the audience change the result? Give an example with two different audiences.
3. When is it better to start a new chat than to keep correcting the old one?