Step 1 — Start With the Goal, Not the Request
Before writing a word, spend 30 seconds clarifying what a successful output actually looks like. Not what you want to ask the AI, but what a perfect response would contain, how long it would be, who would read it, and how it would be used. Starting from the desired output and working backwards prevents the most common failure: prompts that are technically reasonable but strategically wrong. 'I want the AI to write something about our onboarding process' is a request. 'I need a 200-word onboarding welcome email for new enterprise customers, written warmly, that gets them to log in within 48 hours' is a goal. The goal-first approach produces a fundamentally better prompt.
Step 2 — Assign a Role
The first line of your prompt should tell the model who it's acting as. 'Act as a [specific expert with relevant experience]' activates the vocabulary, reasoning style, and domain knowledge of that type of person. Be specific: 'Act as a direct-response copywriter with e-commerce experience' outperforms 'Act as a marketing expert.' The role should match the nature of the task — a technical writing task benefits from a technical writer role, not a marketing expert role, even if both involve writing. If you're unsure what role to use, think about the professional you'd hire to do this task in the real world and describe that person.
Step 3 — Provide Context in 2-4 Sentences
After the role, add the specific background the model needs to produce a useful response for your actual situation rather than a generic one. This includes: who the output is for (audience, knowledge level, relationship to you), what situation it's in (the specific circumstances, constraints, starting point), and what goal it serves (what you're ultimately trying to achieve). Two to four focused sentences of context almost always outperforms a long paragraph — the goal is to eliminate guesswork, not to provide complete backstory. Ask yourself: if I removed this context sentence, would the model's output change significantly? If yes, keep it. If no, cut it.
Step 4 — State the Task Explicitly
Write the action instruction clearly and specifically: what to produce, how long it should be, and what the key requirements are. Use an action verb — write, generate, list, summarize, analyze, compare, rewrite — and then describe the scope of the action. 'Write a 3-section technical blog post' is better than 'write a technical blog post.' 'List exactly 7 specific and distinct actionable tips' is better than 'give me some tips.' If the task has multiple required components, list them as numbered steps: the model handles numbered multi-part instructions much more reliably than long run-on sentences describing multiple requirements.
Step 5 — Add Constraints and Format
End your prompt with the boundaries: what to avoid, how to structure the output, and any technical requirements. Length constraint (max 150 words), format instruction (3 bullet points, no headers), and exclusion rules (no corporate jargon, no passive voice, do not mention pricing) are the three most valuable constraint types for most tasks. Format instructions work best when they describe the exact structure you need rather than just naming a format. 'Three bullet points, each starting with an action verb, max 15 words each' is more precise and more effective than 'bullet points.'
Step 6 — Review, Submit, and Iterate
Before submitting, read your prompt aloud. If any part of it is ambiguous to you, it's ambiguous to the model. Cut anything that doesn't directly help the AI produce the right output — filler, pleasantries, and redundant context all dilute the signal. Submit the prompt and evaluate the output critically. If something is wrong, identify the specific dimension where the output failed (wrong tone? wrong length? wrong content?) and add one targeted fix. Don't rewrite the entire prompt — surgical additions outperform wholesale rewrites for iterative improvement. After 2-3 refinements, save the working prompt as a template for similar future tasks.