Claude for writers: keep feedback separate from drafts

Keep three separate files when you use Claude for writing feedback: your original draft, the conversation, and your revised manuscript. Save feedback as a Word document for review or a Markdown file for your writing app. Record what you used when a publisher or contest asks, and check its AI rules before requesting feedback.

Check the rules before the feedback session

Read the intended publisher’s, contest’s, newsroom’s, or client’s instructions. Check the task you want help with: outlines, creative critique, sentence edits, or drafting. Permission for one doesn’t imply permission for another.

As of November 2026, Amazon KDP’s guidelines require you to report AI-generated book content to KDP, even after substantial editing. They don’t require that declaration for AI-assisted content, such as improvements to writing you created. That distinction belongs to KDP’s policy.

The Bridport Prize’s rules allow limited factual research and checks for spelling or grammar issues. They prohibit AI use for creative development, rewriting, and other creative changes. Permission to identify a typo doesn’t include permission to develop a scene.

Keep the instructions with your project. A log records your process. It isn’t proof of authorship or ownership and doesn’t override submission rules. Ask the recipient about unclear tasks before using AI.

Preserve the draft that Claude is reviewing

Save a versioned original before requesting feedback. Use filenames such as harbor-ch03-v01-original.docx, harbor-ch03-feedback.docx, and harbor-ch03-v02-revision.docx to distinguish the manuscript, conversation, and revision.

Feedback about a missing explanation becomes confusing if you replace the passage and keep only the revised text.

For a permitted critique session:

1. Leave the original draft unchanged.
2. Identify the chapter, scene, or paragraphs in your prompt.
3. Ask one question about the text.
4. Save the response, then record your decisions separately.

Try this illustrative prompt: “Read chapter 3, scene 2 for unclear character motivation. Quote the passages that caused confusion and explain why. Don’t rewrite or propose replacement sentences.”

Read the response before using it. If it supplies rewritten prose anyway, label it as Claude’s suggestion. Record any part you reuse.

Work through the feedback in Word

With DocMyChat installed in desktop Chrome, open the relevant claude.ai chat and choose Export → Word. A .docx file downloads immediately. The Word export guide covers setup and file contents.

The conversation labels turns You and Claude. Check prompts and replies against the chat. Keep the reviewed manuscript and any draft you supplied as an attachment alongside the export. Leave the export unchanged and annotate a working copy if needed.

Open the separate revision file for your manuscript. In Word for Windows, choose Review → Tracking → Track Changes before editing. Microsoft’s instructions explain how Word marks additions and deletions for you to accept or reject.

Work through one feedback point at a time. Mark it “used,” “rejected,” or “needs checking” with a short reason. Put manuscript changes in the revision file. A tracked insertion shows the Word edit. Your AI-use log should identify any Claude wording you retained.

Compare the passage, feedback, and revision

The following passage, feedback, revision, and note are invented examples for a project whose rules allow creative feedback. They are not a real Claude exchange.

• Original draft: “Mara left the key on the table. She had finally decided to leave.”
• Illustrative Claude feedback: “The decision is stated, but the passage doesn’t explain why she leaves now. Clarify the trigger.”
• Illustrative writer’s revision: “Mara set the key beside the letter naming his new fiancée. This time, she left before he could ask her to stay.”

The revision goes in the manuscript, the feedback stays in the conversation record, and the original remains in the earlier draft.

The decision note could say: “Used the critique about missing motivation to revise two sentences in chapter 3. Wrote the replacement sentences separately. No generated wording retained.” This describes the example, not a declaration to copy regardless of what happened.

If you accepted a sentence Claude proposed, record that instead. If you rejected its reading of the scene, note why.

Save one reply for a single critique

For a self-contained critique, use the export icon under that Claude reply and choose Word or Markdown. The single-response guide explains this route.

A reply file leaves out the questions and surrounding discussion. Add a separate note identifying the draft version and passage. Otherwise, “the opening needs more context” may refer to an opening you’ve already replaced.

Keep the full conversation when the requested record includes prompts, follow-ups, or corrections. Don’t substitute one favorable answer for a required account of the session.

Keep a precise record of the help you used

Add an entry to your project’s AI-use log after each relevant session. Record:

• The tool and model, if known.
• The actual conversation date or date range.
• The manuscript version and affected passages.
• What you requested and which suggestions or wording you used.
• The transcript filename and chat URL.
• Checks completed and decisions still outstanding.

“Feedback on chapter 3 motivation” is more useful than “AI-assisted.” Distinguish a critique you acted on from an outline you adopted or replacement prose you retained. Include further help if the session moved beyond its original purpose.

If feedback changed the scene order, record the affected sections even when you kept every sentence. Counting copied phrases alone misses structural choices.

For client copy or journalism, separate factual corrections from style decisions. If a critique questions a product claim or date, note the source you checked and what changed. Leave unresolved factual questions marked for follow-up instead of recording them as completed checks.

The export’s date reflects when you saved it, not necessarily when the conversation happened. Its model appears only if known. Record missing details honestly. Don’t infer an old model name from the one you use today.

Use the AI-use disclosure guide when a publisher, contest, or client requests a statement. Follow the recipient’s wording and placement requirements, and describe only checks you completed.

Keep Markdown feedback beside your Obsidian project

Choose Markdown for feedback in your writing app. Obsidian uses .md files for notes, so move the download into your vault folder.

Keep feedback apart from manuscript drafts. Full-chat exports use You and Claude headings to distinguish speakers. Add your decision note separately instead of rewriting the saved response.

The Claude-to-Obsidian guide covers importing, links, and attachments. Save a new, clearly named file after each session. DocMyChat exports one chat or reply at a time and doesn’t automatically update your vault note.

Questions