Prompting fundamentals that actually work

Why your prompts aren't working

AI models are not mind readers. They respond to exactly what you give them — no more, no less. When you type "write me an email about the meeting", the model genuinely doesn't know: which meeting, to whom, what you want them to do after reading it, what tone is appropriate, or how long it should be.

The output looks plausible. It uses words in the right order. But it isn't actually useful — and that's not the AI's fault. It's a prompting problem, and prompting problems have prompting solutions.

The three-part framework: Task, Context, Format

Every prompt that reliably works contains at least two of these three things. The best prompts contain all three.

• Task — what you want the AI to do. Be specific: not "help me with this" but "rewrite this paragraph in plain English at a Grade 8 reading level."
• Context — why you're asking and what matters. Who is the audience? What constraints exist? What outcome do you need? Context is the biggest lever — we'll dig deeper into it in Lesson 2.
• Format — how you want the output structured. A bullet list? A table? Three options to choose from? A 100-word summary? Specify it, or you'll get whatever the model defaults to.

One task per prompt

A prompt that asks the AI to "summarise this, then rewrite it in a formal tone, then suggest three subject lines, and also point out any factual errors" is asking for four things at once. You will get four average results instead of one excellent one.

When a task has multiple steps, use multiple prompts. Send the summary request first. Review the output. Then ask for the rewrite. You stay in control of quality at every step.

Be specific about what "good" looks like

AI models optimise for satisfying the prompt as stated. If your prompt doesn't define success, the model decides for itself — and its definition may not match yours.

Some of the most effective phrases you can add to any prompt:

• "Give me three versions, each with a different tone."
• "Flag anything you're uncertain about."
• "Do not add caveats or hedging language."
• "Keep it under 150 words."
• "Use bullet points, not prose."

None of these is magic. They work because they narrow the space of acceptable outputs. Narrowing the space is your job as the prompter.

The role instruction

One underused technique: tell the model what role to play. "You are a senior copywriter reviewing a draft for clarity" produces different output than the same request with no role. The model draws on patterns associated with that role — and those patterns tend to be more useful than generic ones.

Good roles to use: "experienced editor", "direct manager giving feedback", "sceptical reader who needs convincing", "technical writer simplifying for a general audience". Avoid vague roles like "helpful assistant" — that's the default.