Why the way you write your prompt changes everything

The most common mistake in early AI use is treating the prompt like a search query. It is not. A prompt is a context — and the difference between an unusable answer and a brilliant one is usually a few sentences of framing that most people skip.

Attention is the whole game

Modern language models do not “look up” answers. They allocate attention across the words you give them and predict what should come next. Every word in your prompt is a vote for what the model should weight more heavily — and silence on a topic is a vote for the model’s prior, which is the average of the internet.

To see why, it helps to know — at a high level — what happens between the moment you press Enter and the moment the model starts writing back. It is the first thing I walk through in every workshop I run, because once people see the mechanism, prompt advice stops sounding like superstition.

The large language models behind ChatGPT, Claude or Gemini are built on an architecture called a Transformer, introduced in a landmark 2017 research paper titled “Attention Is All You Need” (Vaswani et al., Google). The key innovation is the attention mechanism — and you do not need a computer science degree to understand what it does.

How the model actually reads your prompt

When a language model reads your prompt, it does not process words from left to right the way you read a sentence. It evaluates every word in relation to every other word, all at once, asking: given this word, which other words in the prompt matter most for understanding its meaning?

Think of it as a spotlight sweeping across your entire prompt simultaneously. Some words light up strongly in relation to others; some barely register. The model assigns attention weights — numerical scores that determine how much influence each word has on every other word when the model computes its reading of your request.

This is why the same question, phrased differently, produces different answers. A vague prompt gives the model weak, scattered signals: it spreads its attention broadly and produces a generic answer. A clear, specific prompt concentrates the attention weights on what actually matters, and the model’s “understanding” of your intent gets sharper.

This is not a metaphor. It is literally how the technology works. Every word you type is a signal, weighed against all the others, and the model uses the full set to decide what you are asking and how to respond.

Same intent, different prompt, different result

Let me make this concrete with three tasks I see constantly in the teams I work with. Each one shows the same intent expressed as a weak prompt, then as an improved one.

Example 1 — summarising a policy document. The weak version:

Summarise this document.

The model receives almost no signal about what kind of summary you want, for whom, how long, or which aspects to focus on. It will produce a generic overview — often too long, too shallow, or focused on the wrong sections. The improved version:

Summarise this document in 5 bullet points. Focus on the legislative implications for data governance. The audience is non-technical policy staff. Use plain language and avoid jargon.

Now the model knows the format (5 bullets), the focus (data governance), the audience (non-technical) and the tone (plain language). The attention weights concentrate on the sections that match those constraints. The result is shorter, more relevant and immediately usable.

Example 2 — drafting an email. The weak version:

Write an email about the meeting.

Which meeting? To whom? Is this an invitation, a follow-up, a cancellation? The model has to guess all of it — and it will often guess wrong. The improved version:

Draft a follow-up email to the participants of the 15 March coordination meeting on AI governance. Remind them of the three agreed action points: (1) each department nominates a contact person by 29 March, (2) the AI team circulates the draft usage policy by 5 April, (3) the next meeting is scheduled for 22 April. Keep the tone professional but collegial. Maximum 150 words.

Every ambiguity is resolved: the type of email, the audience, the exact content, the tone, the length. The output is draft-ready.

Example 3 — explaining a concept. The weak version:

Explain AI.

The topic is so broad that the model has to make sweeping choices about scope, depth, angle and audience. You might get a textbook introduction, a history of the field or a philosophical essay — none of which may be what you needed. The improved version:

Explain how generative AI differs from traditional rule-based automation, in 200 words, for colleagues who manage document workflows. Use one concrete example comparing how each approach would handle the classification of incoming correspondence.

Precise scope, fixed length, a specific audience, a required illustration. The response is focused and directly applicable to the reader’s work.

In each pair, what changed? Not the model — the prompt.

The four lever moves

• Role. Tell the model who it should be writing as.
• Constraints. Tell the model what not to do, not just what to do.
• Audience. Tell the model who is going to read this and why.
• Examples. Show, don’t tell. One worked example beats a paragraph of instructions.

Those four levers are the version I teach in the room, because four things fit in your head mid-task. When I sit down to write a prompt that actually matters, I unfold them into the five-part structure that prompting practice has converged on: role, context, task, format, constraints. Audience lives inside context and format. Examples ride on top of all five: whenever you can show the model a sample of what good looks like, do it.

1. Role — who is the model in this conversation? Telling the model what hat to wear shapes the vocabulary, depth and framing of the response.

You are an experienced career coach.

Act as a plain-language editor for a general audience.

2. Context — what is the background? Context gives the model the specifics it cannot reliably guess. Without it, the model falls back on generic assumptions — that prior again.

I’m preparing a short briefing for a leadership meeting tomorrow.

This is an email thread with a customer who is unhappy about a delayed delivery.

3. Task — what exactly should the model produce? Use an explicit action verb: summarise, compare, draft, rewrite, brainstorm, critique, translate, outline. A clear verb anchors the output.

Summarise this into 5 bullet points.

Generate 10 headline options.

4. Format — how should the output look? Specify structure, length and tone. Models respond very differently depending on whether you ask for bullets, steps, a template or a short paragraph.

Write it as a 150-word LinkedIn post.

Give me a table with columns: risk, impact, mitigation.

5. Constraints — what should the model avoid or ensure? Constraints are guardrails: what to include, what to exclude, what to double-check, what to prioritise.

Avoid jargon and define any acronym you use.

Do not invent statistics — flag unknowns as assumptions.

Putting it all together

Here is what a complete, well-structured prompt looks like with all five elements in place:

Role: You are a communications specialist writing for a general audience.

Context: I’m writing a short blog post to help people get better results from LLMs like ChatGPT, Claude or Gemini. Many readers are new to prompting and mostly use these tools for everyday work tasks.

Task: Write a short section encouraging readers to try structured prompting on a common task: summarising messy meeting notes.

Format: Maximum 120 words. Use one concrete example. End with a clear call to action. Plain language, professional but approachable tone.

Constraints: Don’t mention company-specific internal tools. Don’t compare models or claim one is “best”. Don’t use fear-based messaging about AI.

This prompt gives the attention mechanism clear targets on every dimension: it knows the role, the situation, the deliverable, the shape and the boundaries. The result will be dramatically more useful than “Write a post about prompting.”

Start simple, then add structure

You do not need to write five-element prompts from day one. Even adding one or two elements will make a noticeable difference. Here is the progression I recommend:

Level 1 — add a task verb. Instead of “AI in healthcare”, write “Explain how AI is used in healthcare diagnostics.”

Level 2 — add audience and format. “Explain how AI is used in healthcare diagnostics, in 200 words, for non-technical policy staff.”

Level 3 — add role and constraints. “You are a healthcare policy analyst. Explain how AI is used in medical diagnostics, in 200 words, for a non-technical audience. Focus on real-world, regulated use cases such as clinical decision support and approved imaging applications. Spell out any acronym the first time you use it, and keep the language plain and concrete.”

Each level builds on the previous one, and the more you practise, the more natural it becomes.

One last consequence of the mechanism — and it is the tip I repeat most often: because every word in the context influences the answer, start a fresh conversation when you change topic, or when a thread gets very long. Old words keep pulling attention long after they have stopped being relevant.

Vague prompts produce vague answers; specific prompts produce specific answers. The difference is not luck — it is structure. Take one task you did today — an email, a summary, a short explanation — rewrite the prompt with the five elements, and compare the two results. That before-and-after is usually all it takes to be convinced.