Most people prompt AI the way they assign work to a person — they describe the task. "Write me an email to follow up with a customer." The result is generic, because the prompt is generic. The single most reliable upgrade in prompt craft is to flip the order: describe the deliverable first, the task second. Tell the AI exactly what the finished thing should look like — the shape, the structure, the constraints, the tone — and the work of producing it almost writes itself.
Task-first vs. output-first
The difference becomes clear once you see the two side by side.
Task-first prompt: "Write me an email to follow up with a customer who didn't respond to my last message."
What you'll get: a generic four-paragraph email with phrases like "I wanted to circle back" and "I just wanted to follow up." It'll be polite. It'll be professionally written. It'll also sound like it was written by an AI to be sent to a stranger, because that's exactly what happened.
Output-first prompt: "Write me a 75-word follow-up email. First sentence: a specific reference to the previous conversation, no generic greeting. Second sentence: a single concrete question that requires a one-line reply. Closing: my name only, no signoff phrase. Tone: warm but direct, like a friend nudging a friend, not a salesperson nudging a lead."
The second prompt produces something usable on the first try. Not because the AI got smarter — because the prompt did the thinking that the task-first version offloaded to the AI.
What "output-first" actually means
An output-first prompt specifies four things the task-first version skips:
1. Shape
How long? What format? Bullet list, numbered list, paragraph, table? Email, memo, summary, outline? The shape is the container the content goes into. Without specifying it, you get the AI's default — usually too long and too formatted.
2. Structure
What goes in which position? First sentence does X. Middle does Y. Last sentence does Z. This is the part that most people skip and that delivers the biggest improvement. AI is excellent at executing structure when you give it one. It produces middling output when you don't.
3. Constraints
What's not in the output. No greetings. No "I hope this finds you well." No bullet points. No questions. The negative space is often more important than the positive instruction — most generic AI output gets generic specifically because of the things it includes by default that you didn't ask for.
4. Voice
Whose voice — and to whose ear? Casual? Formal? Technical? Short? Long? Funny? Serious? The voice is what makes the difference between a paragraph that sounds like you wrote it and one that sounds like AI assistance got involved. Don't say "professional tone." That's not voice. That's a category. Say "like an architect explaining a structural decision to a client — calm, specific, no jargon."
The mental shift
The reason this works is that AI is exceptional at producing things that match a clear specification, and mediocre at deciding what the specification should be. When you give it a task without a deliverable shape, you've handed it the harder half of the work — the half it isn't good at. When you give it the shape first, you've kept that decision for yourself, and reduced the AI's job to execution.
This is true across every kind of output. Writing. Analysis. Code. Lists. Data summaries. The prompts that produce reliably good results are the ones where the writer has done the thinking about what should come out before asking AI to make it come out.
A simple template
If you want a default scaffold for output-first prompting, this one works for almost any task:
Produce a [shape: format and length] that [function: what it accomplishes]. Structure it as: [structure: positions and what goes in each]. Do not include [constraints: explicit exclusions]. Voice: [voice: a specific persona, not a generic adjective].
The five fields in brackets are the entire system. Fill them in honestly and the prompt produces something usable. Skip any of them and the result starts to drift toward the generic AI default.
Examples in practice
Email follow-up
Task-first: "Write a follow-up email to a prospect."
Output-first: "Write a 60-word follow-up email. Open with a specific reference to a detail from our last conversation (assume I'll insert it). Single short paragraph. End with one yes/no question that's answerable in five words. Voice: someone genuinely curious whether the project is still happening, not someone trying to revive a sales process."
Meeting summary
Task-first: "Summarize this meeting transcript."
Output-first: "Produce a one-page summary in three sections: (1) decisions made, (2) open questions, (3) explicit next actions with names attached. No paragraphs of context — only what was decided, what's still pending, and who owes what. Voice: meeting minutes, not narrative recap."
Service page copy
Task-first: "Write me website copy for my drywall repair service."
Output-first: "Write 250 words for a service page on drywall repair. Structure: (1) one-sentence hook describing the customer's problem, (2) two short paragraphs on what the service includes, (3) a three-bullet list of what makes us different, (4) a single closing line that ends with a phone number prompt. Constraints: no marketing clichés ('your trusted partner,' 'second to none'), no superlatives without proof. Voice: a contractor who actually does the work explaining what they do, not a marketing department describing it."
When task-first is enough
For quick, low-stakes asks — a one-line summary, a brainstorm, a what-do-you-think — task-first prompts are fine. The output-first habit becomes critical when the result is going to be used externally, when you'd be embarrassed to send the AI's first draft as-is, or when you'll be using the same prompt repeatedly. The minute the deliverable matters, the minute it pays to spend two extra sentences telling AI what the finished version should look like.
The investment is small. The return is measurable from the first draft.