Clear and direct: a request even a new colleague understands
A model knows only what you write. Write to it as you would to a capable new colleague who knows nothing about your work: what you want, for whom, in what form, and why. It is the simplest and most useful technique there is.
01What you will learn
- Why a model "doesn't guess" — and how to think of it as a new colleague with no context.
- The colleague test: the fastest way to check whether your request is clear.
- The six questions every clear request answers.
- Why one reason does more work than ten bans.
02Before you start
- Access to a chat model: Claude 🌐 global model (the free plan is enough) or a local model 🔒 local.
- One real task from your work this week — a letter, a post, a description, a summary.
- 15 minutes and the willingness to try the same task twice.
03Steps
-
Picture a new colleague
A model is like a very capable person on their first day. It writes well and thinks fast — but knows nothing about your company, your client, yesterday's conversation or what seems "obvious" to you. It knows only what your request says. If something is missing it does not ask — it guesses, and it guesses the average, most likely way.
That is why "Write a post about the event" returns generic, smooth and useless text. Not because the model is weak, but because nobody told it which event, for whom, or why.
-
The colleague test
Before you send a request, read it as if you were handing it to a colleague who missed the meeting. Could they do the job from this text alone, without asking you anything?
✅The ruleIf a colleague with minimal context would be confused by the request, the model will be too. Add whatever you would explain out loud. -
The six questions
A clear request answers six questions. Not always all of them — but check which ones are missing.
Question Example What? a Facebook post for a free seminar For whom? accountants over 40, sceptical about AI How long? up to 80 words What tone — and why? calm and businesslike, because this audience doesn't trust advertising promises In what form? one paragraph, no emoji, no hashtags What does the reader do next? one call to action: sign up via the link Compare the two requests:
UnclearWrite a post about the event.Clear and directWrite a Facebook post for a free seminar "AI for accountants", 24 October, 18:00, Varna. Audience: accountants over 40, sceptical about AI. Up to 80 words, one paragraph, no emoji or hashtags. Tone: calm and businesslike — this audience doesn't trust advertising promises. End with one call to action: sign up via the link.The second request is not "cleverer". It simply says what you would tell a colleague.
-
Say why, not only what
A bare ban works worse than a ban with a reason. When it knows why, the model applies the rule even in cases you did not foresee.
Rule only Rule with a reason Do NOT use ellipses. The text will be read aloud by a voice assistant, so no ellipses — it cannot pronounce them. Keep it short! People will read it on a phone between two clients — three sentences at most. 💡No shoutingCapitals and "MANDATORY!!!" don't help current models — they already follow instructions literally, and pressure makes them overdo it. One calm sentence with a reason is stronger. -
The verb decides
Current models follow instructions more literally. Ask "Can you suggest changes to the text?" and you get suggestions. Write "Change the text so that…" and you get changed text. If you want something specific — no preamble, one name only, exactly three options — say so. The model does not read minds.
⚠️Trap: the polite question"Could you…" sounds polite, but it is a question, not a task. Politeness is fine — just put in the verb for the action: "Rewrite…", "Shorten…", "Put into a table…". -
Practice: your task, twice
Take your task from this week. First write it the way you would have without this lesson. Then write it by the six questions, with a reason for the tone. Compare the two answers. Here is a template to copy and tasks by profession:
Template · clear requestTask: [exactly what to do] For: [who will read or use it] Length: [how much] Tone: [which] — because [reason] Form: [paragraph, table, list, email…] Next step for the reader: [one step] Data to use: [facts, dates, prices — these only]Profession Exercise task Photographer a website image description (alt text) — up to 15 words, for people who can't see the photo Designer a 5-line logo brief — for whom, what feeling, what to avoid and why Marketing an email to past clients — 3 sentences, one link, no pressure Engineer a code-review request — what to check and what not to touch Lawyer a contract summary for a client with no legal training — 5 points, plain language Accountant a reminder to a client about missing documents — polite, with a deadline and a list
04Check
1. What is the fastest way to check whether your request is clear?
2. Why does a rule with a reason work better than a bare ban?
3. You wrote "Can you suggest changes to the text?". What will you most likely get?
4. What is missing from "Write a post about the event"?
05What's next
06Sources
- Anthropic · Prompting best practices — "Be clear and direct" and "Add context"; the living reference for current models.
- Anthropic · Prompt engineering overview — where to start.
- Anthropic · Interactive prompt engineering tutorial, chapter 2 — the original course (written for the 2024 models). The examples in this lesson are our own.
- Anthropic Academy · Claude 101 — a free course with a certificate.