A Guide to Writing a Clear Prompt

Why a prompt deserves ten minutes of your time

Most frustration with AI tools does not come from a weak tool. It comes from a short prompt that leaves the model to guess everything: who is writing, for whom, why, and at what length. The result is text that is "correct" but generic, and you spend a quarter of an hour fixing it by hand. A clear prompt flips the equation: two extra minutes of writing save long rounds of correction.

This guide does not teach "magic words". It teaches a stable way of thinking that works with any chat tool, today and after the tools change: give the model what any new colleague would need to do the job well.

What happens when you write a prompt

Three facts are enough to understand why the method below works:

• The model is not a search engine. It produces text that sounds convincing and may invent a number, a name or a reference. Every number, name and source gets checked before use.
• Whatever you leave out gets filled with the average. No audience means writing for everyone. No tone means a neutral tone. A generic prompt produces a generic result.
• Long conversations blur the context. As a conversation grows, the model may "forget" early details or mix different tasks. A new task deserves a new conversation.

The method: the five-part framework and what completes it

At the heart of the method is the five-part framework: role, context, task, constraints, format. We add three elements that make a prompt ready for real work: the objective and audience before you write, the inputs the model works from, and the quality criteria you will judge the result by, followed by iteration. You do not need every element every time; a quick request needs three, an important task deserves all eight.

1. Objective: start from the use, not the request

Before writing a word, answer: what decision or action comes after this result? "I want a post" is not an objective. "I want a post that gets neighbourhood customers to try the new bread this week" is. The objective shapes everything after it: length, tone and the kind of call to action. If you cannot write the objective in one sentence, the problem is not the tool yet; you have not decided what you want.

2. Audience: who is the final output for?

A summary for the manager is not a summary for the team, and a message to a long-standing client is not a letter to a public body. Describe the reader in one line: rough age or position, what they already know, what they care about. For example: "owners of small shops in Tripoli who do not know technical terms and care about time and cost".

3. Role: who is the model?

The role steers vocabulary and depth. "You are a meticulous copy editor" produces something different from "you are a social media copywriter". Make the role specific and tied to the task, without inflating it: phrases like "the best expert in the world" add no information.

4. Context: what do you know that the model does not?

The model does not know your project or what happened yesterday. In two or three sentences: who you are, what the project is, where things stand now and what has already been tried. Good context is short and specific; do not paste the company's whole history, only what affects this task.

5. Inputs: give it the material instead of letting it invent

If the task is summarising minutes, rewriting a text or analysing a table, paste the material itself and separate it clearly from the instructions, for instance between markers such as """ or under a heading "Text:". Then say explicitly: "Rely only on the material provided; if something is missing, write: not stated." That one line cuts invention considerably.

Before pasting, remove names, phone numbers and any personal or confidential data about clients, patients or staff, unless your tool is an organisational one approved for that use.

6. Task: one clear verb

Write, summarise, analyse, compare, propose, rewrite, extract. One verb per prompt where possible. If you need several, put them as numbered steps inside the prompt or split them across consecutive prompts. "Write a full marketing plan, design posts for it and analyse competitors" is three tasks that deserve three prompts.

7. Constraints and format: the edges of the field and the shape of the ball

Constraints are what the model must respect: length (words or sentences), language or dialect, tone, what must be included and what must be avoided (promises, exaggeration, foreign jargon, numbers not in the material). Format is the shape of the output: a table with named columns, bullet points, an email with a subject line, three alternatives, numbered steps. Name the columns when you ask for a table; it saves you rearranging later.

8. Quality criteria: how will you know the result is good?

This is the most neglected element and the one with the most impact. Write two or three criteria you will use to judge the result, and put them inside the prompt: "the reader must understand the offer from the first sentence", "every number must come from the attached material", "no sentence longer than 20 words". You can also ask the model to check its output against these criteria before delivering it, or to say where it could not meet them.

9. Iteration: the first prompt is a draft

Do not expect perfection on the first try, and do not start from scratch after every disappointment. Improve the result with short, specific follow-ups: "cut it by half", "make the tone friendly but professional", "give me three substantially different alternatives", "the second paragraph is too generic; add an example from our context". Changing one element at a time teaches you what actually moves the result.

Four techniques that lift quality immediately

• Ask me before you start: end the prompt with "If anything is unclear, ask me up to five questions before writing." Very useful when you do not know what is missing.
• One example beats a paragraph of explanation: if you want a particular voice, paste a paragraph of your own writing and say "match this style without copying its phrases".
• Self-check: after the result, ask "Review your answer: which claims are you unsure of? Mark them [unverified]." It does not replace your own checking, but it points to the risky spots.
• The strict critic: "Review the text below and list three weaknesses, most serious first, with a specific fix for each. Do not flatter." Use it on your own writing too.

The prompt canvas: blank template

Copy the template, fill the brackets and delete what you do not need. Save the completed versions that worked in one file; over time it becomes your personal library.

Objective: I will use the result to [the next decision or action].
Audience: the output is for [who, what they know, what they care about].

You are [role: specific expertise tied to the task].
Context: [two or three sentences: who we are, the project, where things
stand, what has been tried].

Inputs:
"""
[paste the text, data or notes here after removing sensitive data]
"""
Rely only on the inputs above. If something is missing, write: not stated.

Task: [one clear verb: write / summarise / analyse / propose / extract].

Constraints:
- Length: [number of words or sentences]
- Language and tone: [plain / formal / friendly / dialect]
- Must include: [...]
- Avoid: [promises, superlatives, numbers not in the inputs, ...]

Format: [table with columns: ... / bullets / email with subject / 3 options].

Quality criteria: before delivering, check that [criterion 1],
[criterion 2] and [criterion 3].
If anything is unclear, ask me one question before you start.

Three examples: from weak to better

Example one: an apology message to a patient

Example two: a post for a new product

Example three: summarising a messy meeting

Ready-made follow-up prompts

After the first result, use one of these short prompts instead of rewriting everything:

When is a short prompt enough?

Not every request needs the full canvas. A quick question ("what is the difference between vision and mission in a business plan?") or a small edit ("fix spelling mistakes only") needs one line. Use the full canvas when you will publish or send the result, base a decision on it, or repeat the task every week. In that last case, save the prompt as a template; it is one of the best investments of your time.

Limits and responsibility

• The model suggests; you are responsible. Every output is reviewed by a person before it reaches another person.
• Do not put personal or confidential data into public tools; anonymise first or use a tool your organisation approves.
• Every number, reference or quotation gets its source opened and checked, ideally against a second independent source for sensitive figures.
• Tool capabilities change constantly (this guide was written in September 2026), but the principles of clarity here do not depend on any particular tool.

Checklist before you send

• I know what I will do with the result and wrote the objective in one sentence.
• I defined the final audience: who they are and what they care about.
• I gave the model a specific role tied to the task.
• I wrote the context it does not know, without padding.
• I pasted the inputs, separated them from the instructions and removed sensitive data.
• The task is one clear verb or a set of numbered steps.
• I set the length, language, tone and what to avoid.
• I specified the output shape (table with columns, bullets, alternatives...).
• I wrote two or three quality criteria inside the prompt.
• I will review and improve the result before using it, and check every number and source.