How to Write Better AI Prompts: A Practical Framework with Examples

Most disappointing AI answers come from prompts that leave the model guessing: who the answer is for, what format is wanted, and what "good" looks like. You do not need special tricks to fix this. A short checklist covers almost every case.

The five parts of a strong prompt

You will not need all five every time, but checking each one before you send catches most vague requests.

• Task: the action you want, stated with a clear verb — summarize, compare, rewrite, debug, plan.
• Context: the background the model cannot guess — audience, product, prior decisions, relevant data.
• Constraints: length, tone, reading level, tools or libraries to use or avoid, deadlines.
• Format: how the answer should be laid out — bullet list, table, numbered steps, code only, a short email.
• Quality bar: what success looks like, or an example of a good answer.

Example 1: a work email

Before: "Write an email about the delay."

After: "Write a short email (under 120 words) to a client explaining that their website launch moves from 3 March to 10 March because payment testing found a bug. Tone: apologetic but confident. Offer a 15-minute call this week. End with a clear next step."

The second version names the reader, the facts, the length, the tone and the ending. The model no longer has to invent any of them.

Example 2: debugging code

Before: "Why doesn't my code work?"

After: "This Express route returns 500 when the request body is empty. Expected: a 400 with a helpful message. Here is the route and the error log. Explain the cause in two sentences, then show the corrected function only."

Including the expected behaviour, the actual behaviour and the error log removes guesswork, and asking for the corrected function only keeps the answer focused.

Example 3: learning a topic

Before: "Explain vector databases."

After: "Explain vector databases to a backend developer who knows SQL but not machine learning. Use one concrete example from e-commerce search. Then give a three-row table comparing when to use a vector database, full-text search, or plain SQL."

Techniques that help in longer tasks

• Ask for a plan first: "Outline your approach before writing anything" lets you correct direction cheaply.
• Request clarifying questions: "Ask up to three questions if anything is ambiguous" prevents confident answers built on wrong assumptions.
• Give examples: one or two samples of the style or structure you want are often clearer than a description.
• Separate instructions from material: put pasted text inside clear markers so the model does not confuse your data with your instructions.
• Iterate deliberately: when an answer is close, say exactly what to change rather than asking it to "try again".

Mistakes that quietly hurt results

• Stacking unrelated tasks in one prompt — split them so each gets full attention.
• Asking for opinions without criteria — say what to optimize for, such as speed, cost or readability.
• Forgetting the audience — an explanation for a beginner and for an expert should look very different.
• Pasting sensitive data — strip passwords, keys and personal details first.

Speeding this up

Once the five-part structure becomes a habit, prompts take seconds to write. If you want help with the first draft, the ContextSwitch Improve Prompt button rewrites a rough prompt into a clearer, structured version inside the chat box on supported assistants, without answering it, so you can review and edit before sending.