How to Write AI Prompts That Produce Useful First Drafts
ost people talk to AI like it’s a smart friend and then wonder why the drafts are vague, off-tone, or useless. Treat your prompt like a tiny spec instead, and you’ll start getting first drafts you can actually ship with light edits.
Loop For Better AI Drafts
- Know your starting levelunderstand
- Minimal AI mental modelapply
- Use 5-part prompt patterndraft
- Write status update emailassess
- Judge good vs poor draftedit
- Debug and fix weak draftsgeneralise
- Extend loop, avoid mistakesrepeat
Table of Contents
- What you’ll be able to do· 1 min
- 1. Know your starting level· 1 min
- 2. A minimal mental model of what the AI is doing· 1 min
- 3. The smallest useful prompt pattern for first drafts· 1 min
- 4. Your first attempt: a status update email· 1 min
- 5. Good vs poor first drafts: what to look for· 1 min
- 6. Rewrite the prompt using the 5-part pattern· 1 min
- 7. Debugging the draft: read like an editor, not a fan· 1 min
- 8. Fixing a weak draft: targeted retry prompts· 1 min
- 9. Extending the loop to other first drafts· 1 min
Cheatsheet: fast prompts for usable first drafts
⚡ 5-part prompt skeleton
Use this structure for most first drafts:
You are a [role].
Your task: [goal – 1–2 sentences].
Audience: [who, what they care about].
Constraints: [2–5 items: length, tone, structure].
Inputs (use these, don’t invent new facts):
- [bullet 1]
- [bullet 2]
Keep prompts in the 60–200 word range for everyday tasks. Shorter than ~40 words is usually too vague; longer than ~300 words often adds noise unless you’re pasting long inputs.
🎯 Quick quality check for first drafts
After the model responds, do a 30–60 second scan:
- Relevance: Are at least 80% of sentences grounded in your inputs? If not, add more detail and retry.
- Tone: Would you send this as-is to the audience? If you’d be embarrassed, specify tone (“direct and informal”, “formal but concise”) and retry.
- Length: If it’s >30% longer than you want, add a word-range constraint and retry.
- Accuracy: Any invented specifics? If yes, tell the model explicitly: “Do not add details not present in my bullets.”
📋 Outline-first vs full-draft-first
Use this rule of thumb:
- If the task is over 800–1,000 words (report, article, strategy doc), ask for an outline first. Then approve or edit it and ask for a draft from that outline.
- If the task is under 500 words (emails, short intros, blurbs), go straight to a full draft using the 5-part pattern.
- If you’re unsure of structure, always start with: “Propose a clear outline for this audience and goal, then wait for my confirmation before drafting.”
🔧 Temperature, length, and settings
If your tool exposes settings:
- Temperature: 0.1–0.4 for business emails, summaries, docs where consistency matters. 0.5–0.7 for creative intros or brainstorming.
- Max length: For emails, 300–600 tokens is plenty; for 1-page docs, 700–1,200 tokens. Too-high limits can encourage rambling.
- If drafts are repetitive across runs, raise temperature by 0.1–0.2. If they’re wild or incoherent, lower it by 0.2 and tighten constraints.
⏱️ Rescue prompts for bad drafts
When a draft misses badly, try one of these instead of starting from zero:
- Tone fix: “Keep all content, but rewrite in a more [casual/direct/formal] tone suitable for [audience]. 150–200 words.”
- Shorten: “Condense this into a version half as long, preserving all key decisions, dates, and asks.”
- De-genericize: “Replace generic phrases like ‘significant progress’ with concrete details from my bullets. If you lack a detail, ask me a question instead of inventing it.”
- Fact-safety: “Only use facts that appear in my input bullets. If something is missing, write a placeholder like [DETAIL NEEDED] instead of guessing.”
Most people talk to AI like it’s a smart friend and then wonder why the drafts are vague, off-tone, or useless. Treat your prompt like a tiny spec instead, and you’ll start getting first drafts you can actually ship with light edits.
What you’ll be able to do
- Use a simple 5-part pattern to write prompts that produce usable first drafts, not fluff.
- Run one real task — a status email — through a full attempt → feedback → retry loop.
- Judge AI drafts quickly and know whether to fix the prompt, feed more input, or just edit by hand.
1. Know your starting level
You don’t need prompt-engineering jargon to get useful first drafts. You do need to know where you’re starting from.
You’re likely in one of these buckets:
- Level 0 – Chatting, not specifying. You type “write an email about project X” and accept whatever comes out, or give up when it feels wrong.
- Level 1 – Some structure, inconsistent results. You sometimes mention tone or audience, but don’t give concrete inputs or constraints.
- Level 2 – Almost there. You give context and paste inputs, but you rarely iterate on the prompt after seeing the first draft.
If you recognize yourself in Level 0 or 1, this guide is for you. We’ll move you to Level 2.5: you can reliably get a workable first draft in 1–2 attempts for everyday tasks.
For the rest of this article, assume you have access to any mainstream LLM (Claude Opus, Claude Sonnet, GPT-4.1, or whatever your tool exposes). The patterns are the same across models.
2. A minimal mental model of what the AI is doing
You don’t need to understand transformers to write good prompts, but you should know what the model thinks it’s doing.
It’s not planning or researching. It’s predicting the most likely next token given your text and its training. Your prompt is just more context for that prediction.
That means:
- If your prompt is vague, the model fills gaps with generic patterns it has seen.
- If your prompt is specific, with real inputs and constraints, the model tends to stay inside those rails.
Think of a prompt as a small spec you hand to a very fast but slightly forgetful junior. The more you pin down audience, purpose, and raw material, the more the draft feels like something you might have written on a good day. When the draft is off, assume your spec was underspecified before you assume the model is “bad.
For first drafts, we’ll use that mental model to narrow the space the model can wander in, without turning your prompt into a wall of text.
3. The smallest useful prompt pattern for first drafts
For most everyday writing, you can get far with a compact pattern:
- Role
- Who the model is pretending to be (changes style and level of detail).
- Goal
- What concrete thing you want (email, outline, 500-word article intro, etc.).
- Audience
- Who will read this and what they care about.
- Constraints
- Length, tone, structure, must-include points, must-avoid items.
- Inputs
- Your actual raw material: bullets, notes, or a rough draft.
You don’t have to label these out loud, but they should all be present.
Example skeleton (keep this handy):
You are a [role].
Your task: [goal].
Audience: [who they are, what they care about].
Constraints: [length, tone, format].
Use this input and cover all its points: [paste bullets / notes].
We’ll use this skeleton in a real task next.
4. Your first attempt: a status update email
Pick a real, low-stakes task you probably do often: a short status update email about a project.
Step 1: Do it the usual (weak) way
Open your AI tool and paste something like:
Write an email updating my manager on our website redesign project.
Ask the model to respond. Don’t help it further. This is the baseline many people live with.
Step 2: Read what you got
Skim the output and notice:
You don’t need to fix anything yet. Just notice what feels off or generic.
5. Good vs poor first drafts: what to look for
Before we improve the prompt, get sharper on what “good enough first draft” means.
Use this as a quick comparison while looking at your email:
| Aspect | Good first draft | Poor first draft |
|---|---|---|
| Relevance | Mentions specific milestones and blockers you recognize | Generic statements that could describe any project |
| Tone | Matches how you normally talk to your manager | Too formal (“Dear Sir or Madam”) or weirdly chatty |
| Length | 5–12 sentences, easy to scan | Multi-screen essay or a 2-line non-update |
| Structure | Clear opening, bullets or short paragraphs, clear next steps | Rambling paragraph, no clear summary or ask |
| Accuracy | Only uses details you supplied or are obviously true | Invents dates, decisions, or approvals you never mentioned |
If your baseline email lands mostly in the right-hand column, that’s not a model problem. It’s a prompt spec problem.
We’ll fix that now.
6. Rewrite the prompt using the 5-part pattern
Take 2–3 minutes to gather actual inputs about your project. For example:
Now structure a better prompt.
You are an experienced product manager.
Your task: Draft a concise status-update email to my manager about our website redesign project.
Audience: My direct manager, who is busy and wants clear milestones, risks, and next steps.
Constraints: 150–220 words, neutral-professional tone, short intro, then bullets under "Done", "In progress", and "Risks / blockers".
Inputs (turn these into the update, don’t invent extra details):
- Launch target: June 30
- Done: new homepage mobile layout implemented and reviewed
- In progress: QA on checkout flow redesign, expected complete by May 20
- Blocker: waiting on legal approval for new terms & conditions page
End with a one-line ask if I need decisions or support.
Run this and compare it with the earlier email.
You should see a draft that sounds closer to how you’d write on a focused day. If it’s still far off, we’ll debug in the next section.
7. Debugging the draft: read like an editor, not a fan
Now you have two drafts: the vague-prompt one and the structured-prompt one.
Read the better one with an editor’s eye:
You’re not judging the model’s intelligence; you’re collecting feedback signals about the prompt:
Capture these in 1–3 short notes. Those notes are the basis for your retry prompt.
8. Fixing a weak draft: targeted retry prompts
Instead of rewriting your entire prompt from scratch, give the model a surgical correction.
Here’s a simple pattern you can reuse:
Here is the draft you wrote:
[PASTE DRAFT]
Problems:
1) [tone issue]
2) [missing details]
3) [length/structure issue]
Rewrite the email fixing only these problems. Keep all accurate project details.
Example filled in for our status email:
Here is the draft you wrote:
[PASTE]
Problems:
1) Tone is too formal; my manager and I usually write in short, direct sentences.
2) It doesn’t explain what the mobile homepage work changes for users.
3) It’s a bit long; I’d like something that fits in ~150 words.
Rewrite the email fixing only these problems. Keep all dates and project facts the same.
Run this. The new draft should feel materially closer to something you’d send.
If it doesn’t, that’s a strong signal to stop iterating and just edit the text yourself. The point is a useful first draft, not perfect automation.
9. Extending the loop to other first drafts
Once you’re comfortable with the status email, you can reuse the same loop for other tasks:
- 1
Draft a meeting agenda from bullets of topics
- 2
Turn rough notes into a one-page summary for a stakeholder
- 3Generate three intro paragraphs for a blog post, using your outline.
In each case, lean on the same components:
You don’t need a different magical prompt for each task. Treat the 5-part pattern as your default and tune it slightly for each use.
10. Common mistakes and how to avoid them
A few patterns reliably tank first drafts. Watch for these and adjust early.
Mistake 1: “Do everything” prompts
Example: “Write a strategy doc for our product for next year.” The model has no facts, no constraints, and no audience.
Fix: Narrow the goal and supply inputs: “Turn these bullets into a 1-page strategy summary for my VP…” and paste your thinking.
Mistake 2: No audience or tone
If you don’t name who you’re writing for, you’ll get a generic business voice. That’s rarely what you want.
Fix: Add one sentence on audience and how formal/casual you are together.
Mistake 3: Treating one output as final
People either accept the first draft or throw it away. Both waste potential.
Fix: Plan on one iteration: quick critique, targeted retry. If it’s still off, edit manually and move on.
Mistake 4: Hiding your real thinking
If you’re shy about messy bullets, you force the model to invent structure and content.
Fix: Paste your messy notes. The model is good at cleaning them up; it’s worse at guessing what you think.
Want a more guided way to practise this?
FAQ: Getting better first drafts from AI
❓ How long should my prompt be for a first draft?
Aim for enough words to cover role, goal, audience, constraints, and inputs, but not a wall of prose. For most tasks, that means roughly 60–200 words beyond whatever raw material you paste in. Under ~40 words, the model has to guess too much and you’ll get generic output. Over ~300 words of instructions, the key constraints can get buried, and some models will ignore them. If your prompt feels long, keep all your bullets, but trim repeated adjectives and vague explanations.
🤔 Should I ask for outlines first or full drafts?
Use outlines when you’re unsure about structure or when the final piece is long. If you’re writing anything above ~800–1,000 words, getting an outline first usually saves time because you can correct the shape before you polish sentences. For short items — status emails, 3-paragraph summaries, short blog intros — going straight to a full draft is fine. A practical tactic is to say: “First, propose an outline for this 1-page doc. Wait for my approval. Then write the draft from that outline and my notes.” That gives you a clean checkpoint.
⚠️ How do I stop the AI from making things up?
You can’t eliminate hallucinations completely, but you can greatly reduce them. First, give the model enough concrete input: dates, decisions, names, or at least placeholders you control. Second, add explicit instructions like, “Do not invent facts; if something is missing, write [DETAIL NEEDED] and stop.” Third, after the draft, scan for any specific numbers or claims you didn’t provide, and either delete or correct them. For high-stakes writing, treat the AI as a drafting assistant and always fact-check critical details before sending or publishing.
💡 Which model should I use for drafting?
For everyday first drafts, most frontier models are good enough: GPT-4.1, Claude Sonnet, Anthropic Opus, or the best model your tool exposes. If you have a choice, favor models advertised as strong at writing, reasoning, and following instructions, not just raw benchmark scores. In practice, pick one, learn its quirks, and stick with it for a while; switching constantly wastes time you could spend refining your prompts. If a draft feels consistently off, it’s usually faster to tighten your spec than to hunt for a different model.
🎯 When should I stop prompting and just edit it myself?
Use a simple rule: if you’ve done two targeted retries and the draft is still structurally wrong or far from your voice, stop and edit manually. The goal is to save time, not force automation. AI is best at turning clear inputs into coherent prose; it’s weaker at guessing your preferences from repeated corrections. If the draft is roughly in the right shape and tone, even with small flaws, it’s often faster to spend 3–5 minutes editing than to design the perfect third prompt. Over time, you’ll develop a sense for which tasks are “one-prompt-and-edit” and which justify more iteration.
Bringing it together: prompts as small specs, not spells
You don’t need elaborate templates or magic phrases to get useful first drafts from AI. You need a clear task, real inputs, and a willingness to iterate once.
The 5-part pattern — role, goal, audience, constraints, inputs — turns a vague request into a small spec. Paired with a quick evaluation pass and one targeted retry, it’s enough to make everyday drafting reliably faster instead of more frustrating.
Treat each first draft as a starting point. If it’s close, edit. If it’s far, adjust the spec, not just the adjectives. Over a week of real use, you’ll collect your own tiny set of reusable prompts, tuned to your work and your voice — far more valuable than any one-size-fits-all list.