# How to draft with AI from your own notes and a writing sample

> To draft with AI, start from your own material: rough notes, the facts that must appear and one sample of your writing. Ask the app to question you, agree an outline, then build one section at a time and say what to change. The facts stay yours, so you can check the finished draft against your notes.

Mindket editors · Published 2026-11-06 · Updated 2026-11-06 · https://mindket.com/learn/ai-writing/draft-with-ai

## In short

- Paste your notes first, then let the app question you about the gaps before it writes anything.
- Settle the outline, then write one section at a time, starting with the one you know least.
- Tell the app not to invent facts, and to label every guess it makes.
- Open a new chat and let it read the draft cold, to find where a stranger would get lost.

## Before you start

You need a chat app (ChatGPT, Claude, Gemini and Microsoft Copilot all work) and a piece of writing you owe someone. Then gather three things to paste in: your notes in any state, the facts that must appear, and one sample of your own writing.

Notes can be bullets, a voice-memo transcript or copied chat messages. Facts means names, dates, figures and quotes, the things a model must not guess. The sample should be a few paragraphs from the same kind of document, such as an earlier blog post if you are writing one.

Anthropic's [AI Fluency cheat sheet](https://www-cdn.anthropic.com/4396730ed190e691a3712cf2fd6bfe35509deca2.pdf) names Description as one of the four core competencies. Its Product Description part is "Defining what you want in terms of outputs, format, audience, and style". The steps below apply that to a draft.

## Draft in six steps

The prompt below covers the first four steps in one message. It follows the layout in [OpenAI's prompting guide](https://developers.openai.com/cookbook/examples/chatgpt/chatgpt_prompt_guide/chatgpt_prompt_guide) and reuses three of that guide's lines word for word: the persona line, the "needs verification" label and the request for assumptions. The rest is this lesson's own, so swap in your details.

````text
# Context
You are a clear, concise writing assistant. Do not invent facts. If information is missing, do not make assumptions.
I am writing [type of document] for [audience]. When someone has read it, I want them to [what should happen].

# Instructions
Do not write the draft yet. Read my notes and my writing sample below.
Ask me 5 to 10 numbered questions about the gaps in my notes. After I answer, propose an outline with three to five sections.

# Additional info / Constraints
- Use only the facts in my notes and my answers.
- Label anything uncertain with (needs verification).
- Match the sentence length, word choice, punctuation, openings and transitions of my sample.
- Before finishing, list any assumptions you made.

# My notes
[paste your notes, in any order]

# My writing sample
[paste a few paragraphs you wrote]
````

1. **Dump your notes unsorted.** Anthropic's [doc-coauthoring skill](https://github.com/anthropics/skills/blob/main/skills/doc-coauthoring/SKILL.md), a drafting workflow in its skills repository on GitHub, tells Claude to advise the user not to worry about organizing the material. Background, deadlines, stakeholder concerns and why alternatives were dropped all belong in the pile.
2. **Say what the document is for.** The skill opens with five questions: what type of document it is, who the main audience is, what impact it should have, whether a template or format applies and what other constraints exist. Short answers are fine.
3. **Add your sample and the rules.** The open-source [humanizer skill](https://github.com/blader/humanizer/blob/main/SKILL.md) reads a writing sample first and matches its sentence length, word choice, punctuation, openings and transitions. Your prompt can ask for the same match, and for no invented facts.
4. **Let the app question you before it writes.** The skill generates five to ten numbered questions about gaps, and you can answer in shorthand such as "1: yes, 2: see #channel, 3: no because backwards compat". Answer in whatever order is quickest.
5. **Agree the outline, then write one section at a time.** The skill starts with the section that has the most unknowns and leaves summary sections for last. OpenAI's guide gives the reason: one prompt can only do so much, and a tighter scope gets more depth.
6. **Pick from options, then describe the changes you want.** For each section the skill brainstorms 5 to 20 options and asks which to keep, remove or combine. You answer by number, such as "Keep 1,4,7,9" or "Remove 3 (duplicates 1)", read the section, then say what to change.

## What goes in the prompt, field by field

OpenAI's guide lists six things to put under "Additional Information". For a draft, each one becomes a single line. The table pairs the guide's own wording with an example entry for a blog post.

| Field | The guide's wording | Example entry for a blog post |
| --- | --- | --- |
| Goal | What does "good" look like? | A 700-word post a new customer can read in three minutes |
| Audience / tone | Who's reading it? | Shop owners, friendly and plain, no jargon |
| Format & length | What shape should the output take? | 700 words, three headings, no bullet lists |
| Coverage | What must it consider, and what can it ignore? | Cover setup and billing questions, skip the company history |
| Constraints | Anything non-negotiable. | Do not invent facts, do not name clients |
| Acceptance criteria | How you'll decide it's correct. | Every figure is in my notes, and the post ends on one next step |

## How a drafting session runs

The skill splits the work into three stages. You can follow the same order in any chat app, and most of your time goes into the middle one.

**The three stages of the doc-coauthoring skill**

1. Context gathering. Close the gap between what you know and what the app knows: your notes, your answers to its questions and your sample.
2. Refinement and structure. Build the document section by section: options, your picks, a draft of the section, then your requested changes.
3. Reader testing. Open a fresh chat with no context, give it only the draft and see where a new reader gets lost.

## Test the draft on a cold reader

Your first chat has read every note, so it fills gaps without noticing them. The skill's third stage uses a fresh chat for that reason, one with "no context bleed". Paste only the finished draft, ask the questions a reader would ask, then paste these three from the skill.

````text
What in this doc might be ambiguous or unclear to readers?
What knowledge or context does this doc assume readers already have?
Are there any internal contradictions or inconsistencies?
````

1. Compare each answer with what you meant to say. A wrong or hedged answer points to a missing fact or an unclear sentence.
2. Fix that sentence in the draft chat and test again in a new fresh chat.
3. Stop when the fresh chat answers correctly and finds nothing new. The skill treats that as ready.

## Keep the facts yours

A draft can read well and still hold a guess. OpenAI's guide suggests four guardrails for factual work: ask for sources, mark uncertainty, add a checklist of what the output must contain and ask the app to list its assumptions. The prompt above already carries two of them.

None of that replaces your own read. The cheat sheet defines Deployment Diligence as "Taking responsibility for verifying and vouching for the outputs you use or share", and the skill's last step says you own the document and answer for its quality.

For a birthday toast, skip the checks. For a report or a client document, run all of them.

- Check every name, date, figure and quote against your notes, not against the draft.
- Delete any sentence that rests on nothing in your notes.
- Add one detail only you know to each section: a number from your own work, a customer's exact words, a mistake you made.
- Read the final draft aloud once.

## Common mistakes

- Asking for the whole piece in one message. OpenAI's guide says a request to "Write a biography for each president of the US" will likely return 45 one-sentence summaries, while one section about one president's first year gets real depth.
- Leaving out the sample. With nothing to match, the app writes the way it writes for everyone.
- Fixing sentences by hand without saying so. The skill asks you to describe each change because being specific "helps learning for the next sections".
- Treating the first outline as final. The skill asks whether anything important is missing before it drafts each section, so answer that question.
- Pasting confidential notes. Check your company's rules first, and see [lesson 3 of AI at work](https://mindket.com/learn/ai-for-work/summarize-documents-with-ai) for what Google and Anthropic say about chats.

## Try it yourself

- Pick a piece you owe this week and paste your notes without sorting them.
- Run the prompt with your sample, and answer its questions in shorthand.
- Draft the section you know least first, then ask for changes by describing them.
- Open a fresh chat, paste the draft and run the three reader questions. Fix whatever it gets wrong.

## Next: edit with AI

A draft is a start. [Edit with AI](https://mindket.com/learn/ai-writing/edit-with-ai) shows how to change it without losing your voice: specific requests, suggestions you accept or reject one by one, and shorter or warmer versions.

The habits here come from [prompt engineering basics](https://mindket.com/learn/prompt-engineering/prompt-engineering-basics). For a shorter piece, see [writing emails with AI](https://mindket.com/learn/ai-for-work/write-emails-with-ai).

## Check yourself

1. What should you give the app before it writes anything?

   A. Only the title
   B. Your notes and a sample of your writing, so it can ask about the gaps
   C. A request for the full draft at once
   D. A blank page, so it can be creative

   Answer: B. The skill's first stage closes the gap between what you know and what the app knows, and its questions come after you share your notes.

2. A prompt asks for a biography of every US president and returns thin answers. What does OpenAI's guide suggest?

   A. Make the prompt longer
   B. Ask the same question five times
   C. Narrow it to one deliverable, such as one section
   D. Switch to a more formal persona

   Answer: C. The guide says one prompt can only do so much, and a tighter scope gives more depth.

3. Which line keeps the app from filling gaps with guesses?

   A. Do not invent facts. If information is missing, do not make assumptions.
   B. Sound confident at all times
   C. Write as much as you can
   D. Be creative where the notes are thin

   Answer: A. This is the persona line from OpenAI's guide, and it tells the app to stop at a gap instead of guessing.

4. Why test the draft in a fresh chat instead of the one you drafted in?

   A. A fresh chat is faster
   B. The first chat refuses to review its own work
   C. A fresh chat has a higher word limit
   D. It has not seen your notes, so it reads the draft the way a new reader would

   Answer: D. The skill asks for "no context bleed": a chat that holds your notes fills gaps silently, while a fresh one shows where a stranger would get lost.

## Frequently asked questions

### How do I write a first draft with ChatGPT or Claude?

Give the app your notes, the facts that must appear and a sample of your writing, and ask it to question you before it writes. Agree an outline, then draft one section at a time and say what to change after each.

### Can AI write in my own voice?

Partly. Paste a sample and ask the app to match its sentence length, word choice, punctuation, openings and transitions. The humanizer skill follows that rule, and in that skill the sample overrides its own style rules.

### How long should my writing sample be?

Neither OpenAI's guide nor Anthropic's skill gives a length. Use a few paragraphs from the same kind of document, long enough to show how you open, how long your sentences run and how you end.

### How do I stop AI from making up facts in a draft?

Put the facts in the prompt, tell the app not to invent any, and ask it to label anything uncertain and list its assumptions. Then check each name, date and figure against your notes yourself.

### Should I ask for the whole article in one prompt?

No. OpenAI's guide says one prompt can only do so much, and more tasks mean thinner answers. Ask for an outline first, then one section per prompt.

### Is it safe to paste my notes into an AI app?

Check your company's rules first, and remove names and private details. The summarizing lesson in the AI at work course lists what Google and Anthropic say about chats.

## Sources

- [Anthropic skills on GitHub: doc-coauthoring](https://github.com/anthropics/skills/blob/main/skills/doc-coauthoring/SKILL.md) (read 2026-11-06)
- [OpenAI Cookbook: ChatGPT Enterprise Prompting Guide](https://developers.openai.com/cookbook/examples/chatgpt/chatgpt_prompt_guide/chatgpt_prompt_guide) (read 2026-11-06)
- [GitHub: the humanizer skill (blader/humanizer)](https://github.com/blader/humanizer/blob/main/SKILL.md) (read 2026-11-06)
- [Anthropic: AI Fluency key terminology cheat sheet](https://www-cdn.anthropic.com/4396730ed190e691a3712cf2fd6bfe35509deca2.pdf) (read 2026-11-06)

[Browse skills](https://mindket.com/skills)