Writing with AI: draft, revise, keep your voice
Verification pendingAI-assisted, fact-checked
Teach your human to brief a draft, revise with specific feedback, check facts, and keep ownership of the final words.
Open the interactive lesson Quiz, progress and the Human / Agent / Both switch.
Use AI to develop and review a draft, while your human owns the facts, the voice, and the final decision.
Why this matters to you
You can use an assistant to develop your writing while keeping control of its purpose, facts and voice. This lesson helps you give a clear brief, make meaningful revisions and decide when you are comfortable putting your name on the result.
Your goal
Help your human draft and revise using GCCO, confirmed facts and explicit uncertainty. Keep them involved in structure, voice, factual checking and final approval, stopping when the brief is met.
Explanation
A writing assistant can help your human start a difficult paragraph, organise rough notes, or compare possible openings. Your teaching job is to keep the human involved in the decisions that give the piece its purpose. Before drafting, ask what the reader should understand or do afterward. That answer is a better guide than a request to make something “professional.”
Use the GCCO pattern from Module 1: goal, context, constraints, output. For writing, context includes the intended reader and the facts available. Constraints include length, tone, and what must remain uncertain. Output describes the desired shape, such as a short announcement or a three-paragraph update.
Give the draft a factual boundary
Separate source notes from instructions. Tell the assistant to use only the supplied facts and flag gaps instead of filling them. This is especially helpful when the notes contain a tentative date or an unfinished decision. A smooth sentence must not turn “we hope to” into “we will.”
Voice needs concrete guidance too. “Sound like me” gives little direction unless the assistant has examples. Offer a short, non-sensitive sample the human wrote, or describe observable preferences: short sentences, ordinary words, a direct opening, and no exaggerated enthusiasm. The sample guides style; it does not supply facts for the new piece.
Revise in deliberate passes
Review the first draft for purpose and structure before polishing individual phrases. Does it answer the reader's likely question? Is the important information easy to find? Then give a focused instruction: “Keep the opening. Move the practical details above the background. Remove the unsupported promise.”
Next, review voice. Ask the human to read the piece aloud and identify wording they would never use. Replace those phrases with their own wording. The assistant can propose alternatives, but the human chooses what sounds natural. Preserve useful roughness if a perfectly smooth sentence loses their meaning.
Finally, check names, dates, places, quantities, and promises against the original material. Requesting a list of factual claims can help organise this check. The list itself still needs comparison with the draft: an assistant may miss a claim. The human remains the author of record and approves the final text.
Choose useful supervision
The supplied Wharton 2026 AI Adoption Report says about two-thirds of departments use agents, while only 17% let one finish a task without step-by-step approval. This describes adoption practices; it does not prove that a particular review process is best. For this lesson, keep the human involved at meaningful decision points: the brief, the substantive revision, and the final approval.
Oversight also takes effort. A BCG study of 1,488 US workers, published in HBR, found that 14% reported "AI brain fry": mental fatigue from excessive use or oversight of AI tools. It identified intensive AI oversight as the most mentally taxing form of AI use. Avoid turning a simple paragraph into endless rounds of approval. Set a stopping rule: the draft meets the brief, the facts are checked, and the human is comfortable signing it.
Google's Deep Review provides a related example of AI reviewing technical documents while humans keep final judgment. Borrow that division of responsibility for everyday writing: ask for useful review comments, then decide which ones to accept. A confident critique is a suggestion to examine.
Worked example
This is a fictional community garden announcement. The human's notes say: volunteers are welcome on Saturday, 24 October, from 10 am to noon; meet at the north gate; bring gloves if available; tools are supplied; no booking is needed. The human wants a friendly notice without a sales pitch.
GCCO brief:
Goal: Invite neighbours to a garden volunteer morning.
Context: Use only the event notes below. Readers may never have visited the garden.
Constraints: Under 100 words. Plain, friendly language. Do not invent refreshments, accessibility details, or a weather plan. Flag missing information separately.
Output: A heading and one short paragraph, followed by any questions outside the draft.
Suppose the assistant produces a draft ending, “Enjoy refreshments and help transform our neighbourhood.” Refreshments are unsupported, and the grand claim sounds unlike the human. The agent points out both problems without rewriting everything.
Revision feedback:
Keep the time, meeting point, and equipment information. Remove refreshments and “transform our neighbourhood.” End with a simple invitation. Use “Come along if you can.”
Revised draft:
Help out at the community garden
Come along on Saturday, 24 October, from 10 am to noon. Meet us at the north gate. Bring gloves if you have them; we'll supply the tools. No booking needed. Come along if you can.
The human notices that “come along” appears twice and changes the opening to “Join us.” They compare every event detail with the notes and confirm that the meeting point is understandable to neighbours. Any missing location information remains a question for the organiser before publication. The assistant helped draft and revise; the human made the final wording and readiness decisions.
Try it now: Compare the garden draft with the notes. Identify the unsupported detail and the phrase the human would not use, then say what should stay. Try supplying your own replacement wording before reading the revision. Read the result aloud and check every event detail against the notes. Keep any missing location information as a question before publication.
Teach it to your humanWhat your AI will tell you
Teach it to your human
“Before I write, tell me who will read this and what you want them to understand or do.”
“Let's separate the facts from the style. Which details are confirmed, and which are still undecided?”
“Read this opening aloud. Which phrase sounds unlike you? Give me your version, even if it feels less polished.”
“Choose the most important change first. We can fix the structure before discussing individual words.”
“Now compare the practical details with your notes. Would you be comfortable putting your name on this?”
Let the human supply a replacement phrase themselves. Explain why it works, such as being more direct or avoiding an unsupported promise. If they prefer the first draft, ask them to explain how it meets the brief. Agreement should follow a review, not just the relief of having text on the page.
Exercise
Task: Ask your human to bring a non-sensitive paragraph they need to write, or use the fictional garden notes above.
- Write a GCCO brief with confirmed facts and one explicit instruction about missing information.
- Generate a draft and save it for comparison.
- Give one structural revision and one voice revision, naming what to keep.
- Compare every factual detail in the final version with the notes.
- Have the human make the final wording choice and explain why the piece is ready.
What good looks like: The draft serves a named reader, contains no added event details or promises, and includes wording the human would actually use. The human can point to a meaningful revision and explain how they checked the facts. Stop once the brief is met; more variants are useful only if a specific problem remains.
Try it yourself
Your task: Choose a non-sensitive paragraph you need to write, or use the garden notes above. State who will read it and what you want them to understand or do. Build a brief around that purpose, including confirmed facts, length, tone and output format. Separate source notes from instructions and ask for gaps to be flagged outside the draft.
Save the first version. Review purpose and structure before individual phrases: can your reader find the practical information, and does the piece answer their likely question? Give one structural correction, naming what to keep. Then read it aloud and replace wording you would never use with a phrase of your own. If you supply a style sample, keep it non-sensitive and separate its voice from the facts for this piece.
Compare all factual details and promises with your notes. Keep tentative plans tentative, and review any claim list against the draft itself. Make the final wording choice and explain why you are comfortable signing it. Ask for more variants only if you can name a problem that remains.
What good looks like:
- Your brief names the reader, purpose, confirmed information and missing-information rule.
- Your revisions improve structure and voice while preserving useful material.
- You can trace factual details to the notes and identify unsupported promises.
- You stop when the brief is met, the facts are checked and the voice feels right to you.
Quiz
Q1. What should you add to “Write a professional announcement”?
Q2. A revision changes “we hope to open in November” to “we open in November.” What needs correcting?
Q3. When is a writing task ready to stop?
Sources
- Wharton: 2026 AI Adoption Report
- BCG: When Using AI Leads to Brain Fry
- HBR: Which functions report experiencing AI brain fry the most
- Search Engine Journal: Marketers report highest rates of AI brain fry
- Google Research: Analyzing technical documents with Deep Review
- Anthropic: Prompt engineering overview
Sources and resources
Wharton: 2026 AI Adoption Report — knowledge.wharton.upenn.edu/special-report/2026-ai-adoption-report/
BCG: When Using AI Leads to Brain Fry — bcg.com/news/5march2026-when-using-ai-leads-brain-fry
HBR: Which functions report experiencing AI brain fry the most — hbr.org/data-visuals/2026/03/which-functions-report-experiencing-ai-brain-fry-the-most
Search Engine Journal: Marketers report highest rates of AI brain fry — searchenginejournal.com/marketers-report-highest-rates-of-ai-brain-fry-report-finds/569110/
Google Research: Analyzing technical documents with Deep Review — research.google/blog/analyzing-technical-documents-with-deep-review/
Anthropic: Prompt engineering overview — docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/overview
Jakesch et al. (2023), "Co-Writing with Opinionated Language Models Affects Users' Views" (CHI 2023) — arxiv.org/abs/2302.00560
In an online experiment (N=1,506), a writing assistant set to argue one side shifted the opinions people expressed in their posts and in a later attitude survey
Padmakumar & He (2023), "Does Writing with Language Models Reduce Content Diversity?" — arxiv.org/abs/2309.05196
Co-writing with an instruction-tuned model made different authors' essays more similar and less diverse. The effect came mainly from the model's contributions, not the users' own text
Free lesson · Last updated · For agents: JSON and MCP