How to Write a Prompt That Gets a Usable First Draft

Most prompts fail for the same reason: they name the task and nothing else. This guide sets out the four things a prompt has to state, the order to write them in, how to hand over context without pasting an entire document, and how to tell when the prompt is not the thing that is wrong.

The short answer

A prompt produces a usable first draft when it states four things: the job, the reader, the shape of the output, and what must not happen. Write them in that order, one or two sentences each, then ask for a single section before you ask for the whole piece.

Most prompts state only the job. That is why most first drafts come back as something you could have written faster yourself: correct in subject, wrong in every other dimension, and shaped like the average of all the text ever written on the topic.

The word usable is doing real work there. A usable draft is not a finished one: it is a draft where the structure is right, the register is close, and the work left is editing rather than starting again. That is the realistic target, and a prompt that fits in a short paragraph will reach it.

Everything below is the long version: why the default output disappoints, how to write each of the four parts, how to hand over context without pasting a whole document, how to ask for a revision instead of a replacement, and how to recognise the cases where no prompt will fix the problem.

Why the first draft usually comes back unusable

When a prompt leaves something unstated, the assistant does not stop and ask. It fills the gap with the most common answer, and the most common answer is an average of an enormous amount of published text.

That is the mechanism behind almost every complaint about generated writing. The draft is too long because unspecified length defaults to thorough. It explains things your reader already knows because an unspecified reader defaults to a general audience. It hedges every claim because unspecified confidence defaults to safe. It opens with a paragraph restating the question because that is how an enormous number of articles open.

None of this is the model being unhelpful. Each default is a reasonable guess at an unanswered question, and together they produce text that reads as though it were written for nobody in particular, because it was.

This reframes what a prompt is for. There is no secret phrasing that unlocks better output. A prompt is a list of decisions you are making so that the assistant does not have to make them by averaging, and every sentence you add should remove one default. A sentence that removes no default is costing you attention and adding nothing.

The four things a prompt has to state

These four cover the decisions that cause the most damage when they are left open. Keep them in this order, because each one constrains the next.

The first three are familiar. The fourth is the one almost everyone skips, and it is often the highest-value sentence in the prompt, because an assistant has no way to know which conventions you find intolerable unless you say so.

• Job: the deliverable in one sentence, with the verb that matters. Draft, summarise, outline, rewrite, extract and compare are different jobs and produce different text.
• Reader: who they are, what they already know, and what decision they are trying to make. This single line does more to fix register and length than any instruction about tone.
• Shape: approximate length, how many sections, whether headings are wanted, and whether the output should be prose or structured. State it even when it feels obvious.
• Limits: what must not happen. No invented examples, no statistics, no opening summary, no closing summary, no comparison to a competitor, no claims about what the reader feels.

Step one: name the job in one sentence

Write the deliverable, not the subject area. The difference is the difference between a draft you can edit and a wall of general background.

A topic-shaped prompt asks about a subject and gets an encyclopaedia entry. A job-shaped prompt names an artefact, and the artefact constrains everything that follows: a one-page briefing note for a manager who has to approve a purchase is a completely different object from an explanation of the same subject, even though both are about the same thing.

The verb carries most of the meaning, so choose it deliberately. Summarise means compress what is already there. Draft means produce new prose. Outline means structure without sentences. Rewrite means keep the content and change the surface. Extract means pull out what matches a pattern and add nothing. Assistants follow these verbs fairly reliably, so a vague one is a wasted opportunity.

Also say what the output is for. Purpose resolves a surprising number of ambiguities on its own: a paragraph destined for a proposal and a paragraph destined for an internal note need different levels of caution, and naming the destination settles that without a separate instruction.

Step two: describe the reader, not the subject

This is the step that most improves a draft relative to how little it costs to write, and it is the step most often left out.

Without a reader, the assistant writes for a general audience, and writing for a general audience means explaining everything a little and nothing properly. You get definitions of terms your reader uses daily, and silence on the one point they actually need. With a reader named in a single line, the same prompt produces text pitched at a level you can work with.

Say three things: who they are, what they already know, and what they want from the text. The second is the important one. Naming what the reader already knows is the only reliable way to stop an assistant laying foundations nobody asked for, because it converts a vague instruction about being concise into a concrete rule about what to leave out.

Avoid the role-play shortcut here. Telling an assistant to write as a specialist is an indirect way of describing the reader, and less precise than describing the reader. State the audience and the register follows. State a role and you are hoping the model infers the audience.

Step three: specify the shape of the output

Shape means the physical form of the text: how long, how many parts, what each part contains, and what format each part is in. State it even when it feels too obvious to mention, because obvious to you is unstated to the assistant.

Length comes first, in approximate terms rather than exact ones. Asking for a precise count is unreliable in a way that frustrates people, because these tools do not count words as they write. A range, or a comparison to something familiar such as a short paragraph or a single page, works far better.

Structure comes second. Say whether you want headings, how many sections, and what each section is responsible for. If the output has to slot into an existing document, describe the slot. If you want the assistant to stop after the first section so you can check direction, say so, because the default is to produce the whole thing.

Format comes third: prose, list, table or a mixture, and whether any part needs to be machine-readable. Mixed formats work, but only when you say which part is which.

Step four: say what must not happen

A prompt made only of positive instructions still leaves every house habit in place. The limits section is where you turn those off, and it is where a reusable prompt earns most of its value, because the same handful of habits will otherwise appear in every draft you ever generate.

The habits worth banning by default are easy to list once you have noticed them. Opening paragraphs that restate the question. Closing paragraphs that summarise what the reader just read. Transitions that announce structure instead of carrying an argument. Sentences that give equal weight to a strong point and a weak one. Claims about what the reader thinks or feels. Invented examples presented as observations.

The last one deserves its own line in every prompt. Left unconstrained, these tools produce plausible specifics: a figure, a date, a named study, a customer anecdote. They are generated to fit the sentence, not retrieved from a source, and they are the most expensive thing to miss in an edit, because they read exactly like the parts that are true.

Write the limits as instructions, not preferences. Do not include statistics is followed. Try to avoid too many statistics is not.

How to supply context without pasting the whole document

The instinct when an assistant lacks context is to paste everything. It is the wrong instinct, and not mainly because of length limits.

A long paste dilutes your instructions. The assistant now has a great deal of material and no indication of which parts matter, so it treats the document as the subject and your prompt as a footnote. Drafts produced this way tend to mirror the structure and tone of whatever you pasted, including the parts you were hoping to replace.

Hand over context in three deliberate pieces instead. Give a short statement of what the source document is and what role it plays. Give only the extracts that bear on the specific job, clearly marked as source material. Then state what the assistant should do with them, including whether it may contradict them.

That last point prevents a quiet failure. Paste a paragraph without saying why and the assistant will usually treat it as a model to imitate. Say that it is the current version and too cautious, and the same paragraph becomes something to improve against.

When a document genuinely has to go in whole, name the sections that matter and say the rest is background only.

Ask for a revision, not a rewrite

When a draft is close but wrong in places, the costly mistake is to repeat the whole request with an added complaint. That produces a new draft rather than a corrected one, and new drafts bring back problems you had already fixed.

Ask for a revision instead, and be specific about scope. Name the part that needs to change, say what is wrong with it, say what should replace it, and state that everything else stays as it is. The last clause matters more than it seems, because without it the assistant will rewrite material it was never asked to touch and you will spend the next round restoring it.

Make the diagnosis concrete. Too formal is weaker than this assumes the reader already agrees. Too long is weaker than cut the paragraphs that repeat the opening. A specific diagnosis gives the assistant something to act on; a vague one gives it another chance to average.

If two rounds leave the same problem in place, stop revising. The instruction is competing with something earlier in the thread, and a third attempt usually produces drift rather than progress.

Three mistakes that account for most bad drafts

These three come up more than everything else combined, and all three are easy to fix once named.

The first is asking for everything at once. A request for a full piece with several sections, a particular structure and a specific tone gives the assistant too many competing constraints, and the ones at the end tend to lose.

The second is piling on adjectives instead of constraints. Engaging, professional, compelling and authoritative are not instructions, because they rule nothing out. Replace each with a statement about the reader or a ban on a habit, and the draft changes immediately.

The third is leaving a thread running too long. Context accumulates, including instructions you have since reversed and drafts you have rejected, and the output starts drifting back toward earlier versions.

• Asking for the whole deliverable in one request instead of outline, then section by section.
• Using adjectives about quality where a constraint about the reader or a banned habit would do the work.
• Continuing a thread after you have contradicted yourself, so that both instructions are still live.

When the prompt is not the problem

Some tasks do not get better with better prompting, and recognising them early saves more time than any technique on this page. There are three cases.

The wrong tool. A general assistant asked to work across a large body of your own material will do it badly no matter how the request is phrased, because the material is not available to it. The fix is a tool built to search your documents, not a longer prompt.

The wrong data. If the facts you need are not in the prompt and not something the assistant can look up, no phrasing will produce them. What you will get instead is fluent text with plausible specifics in it, which is worse than an empty page because it looks finished. Supply the facts or accept that the output is structure only.

The wrong job. Some work cannot be delegated because the judgement is the deliverable: deciding what to recommend, choosing what to leave out, taking responsibility for a claim. An assistant can draft the sentences around that judgement but cannot make it, and a prompt that asks it to is asking for a guess dressed as a decision.

When you hit one of these, stop iterating. More prompting is motion without progress.

Check the draft before you use it

Drafting and checking are separate jobs, and the second one does not get easier because the first one went well. Fluent text produced quickly feels verified in a way that it is not, which is exactly why this step gets skipped.

Run the draft through four passes, in this order. Check every specific first: names, dates, figures, quotations, product capabilities, anything that could be right or wrong. Each needs a source you can open. Anything you cannot confirm comes out rather than staying in with a hedge.

Then check the claims. Fluent prose makes weak reasoning sound settled, and a sentence can be factually accurate and still assert more than its evidence supports. Read for the gap between what is shown and what is concluded.

Then check for absences. Generated drafts omit the awkward case, the caveat and the counter-argument, because the averaged version of a topic is tidier than the real one. Ask what a knowledgeable reader would object to, and whether the draft answers it.

Finally read it aloud. Repeated sentence shapes are almost invisible on screen and obvious in the mouth.

The checklist, in order

Work down this list before you send a prompt, and use the lower half before you send the output anywhere.

Name the deliverable and its purpose in one sentence, choosing the verb deliberately. Describe the reader, including what they already know and the decision they are making. State the shape: approximate length, number of sections, what each section does, and whether the output is prose or structured. Write the limits as explicit bans, with invented specifics always among them. Supply context as marked extracts, and say whether the assistant may contradict them. Ask for an outline first, then one section at a time. When something is wrong, ask for a scoped revision with a concrete diagnosis. Abandon the thread after two reversals.

Then, before the draft goes anywhere: verify every specific against a source you can open; read for claims that outrun their evidence; ask what is missing rather than what is wrong; read it aloud.

If it is still unusable after all of that, the problem is likely the tool, the data or the task. How these guides get written and checked is set out in our editorial policy, and who stands behind them is on the about page.

How we chose

This guide is method rather than measurement. It names no tool, quotes no price and reports no output quality, because any of those would be true only of one model on one day, and this page is written to still be correct after the next release. What it describes is the structure of the instruction itself, which has stayed stable across every generation of these tools: a task, a reader, a shape, and a boundary. Follow the steps on whichever assistant you already have open. If a step stops working, our preference is to fix the page rather than defend it.

Frequently asked