Polished first drafts
Produce a real first draft of a memo, report, or letter from a brief and a template — clean enough to revise, not a starting point you have to throw away.
Drafting and revising real documents — memos, reports, letters, proposals — in your voice or someone else's, without leaving the AI-tell phrasing that makes a Claude-written draft obvious to the reader.
Start here
Produce a real first draft of a memo, report, or letter from a brief and a template — clean enough to revise, not a starting point you have to throw away.
Edit existing copy without losing the voice it was written in — yours, your manager's, the company's. Voice preservation is half of document work.
Produce .docx files that open cleanly, print right, match the template, and don't need anyone's manual cleanup before they go out.
Three document moves
The first draft is produced from a real brief. Audience, length, voice, key points, format. Skip the brief and Claude will guess — and the guess is usually generic, hedged, and tonally off.
Revisions are surgical. "Tighten paragraph three. Cut the second sentence. Replace the closing with a question." Section-level edits preserve everything you already approved.
The voice the reader expects is the voice the document needs to have. Yours, your manager's, the company's — never Claude's default, which has a recognizable AI flavor that subtly tells the reader the work was generated.
Where the work happens
The main surface for document work. Cowork opens the file, modifies sections, applies template formatting, and saves a real .docx that opens cleanly in Word.
Main surfaceBefore any prose exists, outline and tone are sorted in chat. What sections, what order, what voice. Discussions live in chat; the document lives in Cowork.
Outline firstThe brand-voice or letterhead template that defines structure, fonts, and formatting. Point Claude at it — don't have Claude invent it.
Real templateDocument work
Audience, length, voice sample reference, the three or four points the document has to make, the format the recipient expects. That's the whole brief — and it earns most of its time back in the first draft.
One real document in the target voice — your last good memo, your manager's last update, the company's published letter — beats a paragraph of describing what you want. Claude learns from examples faster than from adjectives.
For anything longer than a single email, get a numbered outline first. Approve the structure, then generate the prose. Restructuring after the fact is the most expensive editing pass.
"Rewrite the second paragraph to be tighter" beats "rewrite the whole thing." Section-level revision keeps the parts that work and only touches what's broken.
Last move before save-as-final: ask Claude to scan its own draft for AI tells, banned phrases, and voice mismatches. Treat the output as a checklist you read line by line.
Prompt upgrade
Write me a memo on the new appraisal-gap policy.
Document-ready prompt
(In Cowork.) Draft a one-page internal memo on the new appraisal-gap overlay. Audience: GMFS LO and processor teams. Voice: factual, friendly, no corporate filler — match my-voice-sample.docx in this folder. Key points: 1) what changed, 2) effective date, 3) what LOs do differently, 4) who to ask. Use the GMFS-memo-template.docx. Three short sections, fits one printed page.
Best patterns
Rough bullet notes in, polished one-page memo out. Highest-volume document use case for most employees.
Outline pass first, section pass second, voice pass last. Long documents fail because the outline pass got skipped — fix that, and the rest converges fast.
Real letterhead, real voice sample, single audience. Where a Claude-written draft is most likely to pass for human-written if the brief is right.
"Here's the draft. Here's what's wrong with it. Rewrite, preserving everything I didn't flag." Surgical revision is more useful than re-drafts.
Document checks
Draft and revise
The fastest way to improve here is to read your own drafts back the way a careful reader would. The detector in this lesson is a starting point.
You've finished this module when you can produce a first-draft .docx that's 80% shippable, revise in someone else's voice without losing it, and catch your own AI tells before a reader does.
AI-tell detector
. regarding our recent conversation about the upcoming policy changes that will be implemented next quarter. As we discussed, we want to ensure all stakeholders . I would be more than happy to schedule some time to the specifics of how these changes will impact our operations. Please don't hesitate to reach out if you have any questions or concerns. Looking forward to our continued partnership.
Quick note before next quarter's policy change. Want to make sure we're on the same page about how the appraisal-gap overlay will affect our LOs day-to-day. Got 15 minutes this week to walk through the impact on the pipeline? I have Tuesday afternoon open — pick a time that works and I'll send the meeting.
— Dan
None of the AI tells are wrong, exactly. They're just predictable — which is the tell. A reader can feel the cadence of "I hope this email finds you well" before they finish reading it. That's why they signal "AI wrote this" so quickly.
A document with two visible AI tells reads as AI-written to a careful reader — and once a reader notices, they read everything else through that filter. Spotting tells is the cheapest credibility move available.
Before sending anything Claude touched, read it once asking: "would I have written this sentence?" If the answer is no twice in one paragraph, rewrite the paragraph in your voice.