Index a newsletter archive and repurpose writing
Last updated: June 10, 2026. Task: Last updated: June 10, 2026 Somewhere in your ChatGPT history is a paragraph you loved, buried under four hundred messages and a scroll bar. That experience is the whole argument for Claude Code for writers: your drafts become files on your computer, your style guide becomes a file the AI reads automatically at the start of every session, and edits arrive as visible diffs you approve line by line. The chat never has to remember anything, because the folder does. I’m Carl. I run a free course that teaches Claude Code to people who do not code, and the writers who finish it (essayists, novelists, journalists, newsletter people) are its most surprised graduates. Most guides to Claude Code for writing were written by developers, for developers. This one assumes you write for a living and code not at all. Why Claude Code for Writers Starts With Files, Not Chat A file survives; a chat thread decays. Writers who co-write in chat windows all hit the same wall, the one novelist Marilise de Villiers describes: “when conversations got too long, Claude would lose context and I’d have to start fresh… re-establish the voice”. Every long chat session ends with you re-teaching the AI who you are. Claude Code skips the re-teaching. It is Anthropic’s AI agent that runs in a terminal (also a desktop app, also on the web at claude.ai/code), reads and writes the actual files in a folder, and takes instructions in plain English. Point it at drafts/essay.md and it edits that file, on your disk, where any editor can open it. Close the laptop Tuesday, come back Thursday, and everything is exactly where you left it, because none of it ever lived inside a conversation. Writers doing this are no fringe case. Anthropic’s own usage research found that arts and media work, mostly copyediting and the refinement of fiction, is the third-largest category of Claude usage, and it grew through late 2025 while coding’s share shrank. Tech writer Robert Matsuoka, who moved his entire article pipeline over, estimates that killing copy-paste alone saves him about 30 minutes per article. His framing is the cleanest I’ve seen: generate with AI, edit with editors, stop forcing one tool to do both. Claude Code handles the files; he polishes in Obsidian. If the word “terminal” makes you want to close this tab, the non-developers guide builds the mental model gently, and the install takes about 15 minutes. The Edit Pass: Claude Code Editing With a Style Guide in CLAUDE.md One file changes everything about AI editing: CLAUDE.md. It sits in your project folder, and Claude Code reads it automatically at the start of every session. Write your style rules there once and they ride along forever, no re-pasting, no “as I mentioned earlier.” Type /init and Claude drafts the file for you; the course covers the details in project memory. Two specifics from Anthropic’s docs are worth respecting. Keep the file under 200 lines, because long files reduce adherence. And remember it is context rather than law: strong steering, never enforcement, which is exactly why the diff review step below exists. 1Draft it yourself, uglyWrite the messy first version in any editor. The thinking stays yours. 2Drop the file in your project folderClaude Code reads it directly with @filename. Nothing gets pasted. 3Critique in plan modePress Shift+Tab to cycle to plan mode, which makes Claude read-only. Ask for the five biggest problems, ranked, before anything changes. 4Run one bounded edit passEdit, do not rewrite: one pass for clarity only, with your CLAUDE.md rules steering every change. 5Review the diff line by lineReject anything that does not sound like you. Rejections are the system working. 6Read it aloud and shipFinal wording is a human decision, and no amount of tooling changes that. The diff is the feature. When Claude Code edits a file, it shows precisely which lines changed, old version next to new, and you accept or reject each one. A chat rewrite forces you to re-read the whole passage hunting for what moved; a diff makes drift visible at a glance. The prompt that drives a good pass is short: The @ syntax points Claude at a specific file, and it is the single most useful habit to build first; working with files is the course lesson on it. Separating critique from editing matters too. Plan mode gives you a reader who can find every problem and change nothing, which is rarer in real life than it should be. AI Will Flatten Your Voice If You Let It; tools: Claude Code. Human role: The original post does not say how a person stayed involved.; limitations: Claude cannot reliably count, so it may miss repeated phrases. Long generation sessions degrade quality. CLAUDE.md should be under 200 lines for better adherence. It is context, not law, requiring human diff review. Scrivener and Word require export to markdown. Proposed advice: Start bounded to avoid invented tags and chaos. Verify quotes against transcripts..
Tools
Published on Agent Work Atlas:
Overview
Last updated: June 10, 2026
Workflow
Last updated: June 10, 2026 Somewhere in your ChatGPT history is a paragraph you loved, buried under four hundred messages and a scroll bar. That experience is the whole argument for Claude Code for writers: your drafts become files on your computer, your style guide becomes a file the AI reads automatically at the start of every session, and edits arrive as visible diffs you approve line by line. The chat never has to remember anything, because the folder does. I’m Carl. I run a free course that teaches Claude Code to people who do not code, and the writers who finish it (essayists, novelists, journalists, newsletter people) are its most surprised graduates. Most guides to Claude Code for writing were written by developers, for developers. This one assumes you write for a living and code not at all. Why Claude Code for Writers Starts With Files, Not Chat A file survives; a chat thread decays. Writers who co-write in chat windows all hit the same wall, the one novelist Marilise de Villiers describes: “when conversations got too long, Claude would lose context and I’d have to start fresh… re-establish the voice”. Every long chat session ends with you re-teaching the AI who you are. Claude Code skips the re-teaching. It is Anthropic’s AI agent that runs in a terminal (also a desktop app, also on the web at claude.ai/code), reads and writes the actual files in a folder, and takes instructions in plain English. Point it at drafts/essay.md and it edits that file, on your disk, where any editor can open it. Close the laptop Tuesday, come back Thursday, and everything is exactly where you left it, because none of it ever lived inside a conversation. Writers doing this are no fringe case. Anthropic’s own usage research found that arts and media work, mostly copyediting and the refinement of fiction, is the third-largest category of Claude usage, and it grew through late 2025 while coding’s share shrank. Tech writer Robert Matsuoka, who moved his entire article pipeline over, estimates that killing copy-paste alone saves him about 30 minutes per article. His framing is the cleanest I’ve seen: generate with AI, edit with editors, stop forcing one tool to do both. Claude Code handles the files; he polishes in Obsidian. If the word “terminal” makes you want to close this tab, the non-developers guide builds the mental model gently, and the install takes about 15 minutes. The Edit Pass: Claude Code Editing With a Style Guide in CLAUDE.md One file changes everything about AI editing: CLAUDE.md. It sits in your project folder, and Claude Code reads it automatically at the start of every session. Write your style rules there once and they ride along forever, no re-pasting, no “as I mentioned earlier.” Type /init and Claude drafts the file for you; the course covers the details in project memory. Two specifics from Anthropic’s docs are worth respecting. Keep the file under 200 lines, because long files reduce adherence. And remember it is context rather than law: strong steering, never enforcement, which is exactly why the diff review step below exists. 1Draft it yourself, uglyWrite the messy first version in any editor. The thinking stays yours. 2Drop the file in your project folderClaude Code reads it directly with @filename. Nothing gets pasted. 3Critique in plan modePress Shift+Tab to cycle to plan mode, which makes Claude read-only. Ask for the five biggest problems, ranked, before anything changes. 4Run one bounded edit passEdit, do not rewrite: one pass for clarity only, with your CLAUDE.md rules steering every change. 5Review the diff line by lineReject anything that does not sound like you. Rejections are the system working. 6Read it aloud and shipFinal wording is a human decision, and no amount of tooling changes that. The diff is the feature. When Claude Code edits a file, it shows precisely which lines changed, old version next to new, and you accept or reject each one. A chat rewrite forces you to re-read the whole passage hunting for what moved; a diff makes drift visible at a glance. The prompt that drives a good pass is short: The @ syntax points Claude at a specific file, and it is the single most useful habit to build first; working with files is the course lesson on it. Separating critique from editing matters too. Plan mode gives you a reader who can find every problem and change nothing, which is rarer in real life than it should be. AI Will Flatten Your Voice If You Let It
Reported result
Casey Newton indexed 818 newsletter editions in under 30 minutes. Writers report that vague instructions fail, while binary rules in CLAUDE.md stick. Unsupervised AI editing can flatten voice, replacing jokes with summaries. Repurposing saves time, with one writer estimating 45 minutes per article. Success depends on human review of diffs and adherence to style rules.
Human role
The original post does not say how a person stayed involved.
Limitations
Claude cannot reliably count, so it may miss repeated phrases. Long generation sessions degrade quality. CLAUDE.md should be under 200 lines for better adherence. It is context, not law, requiring human diff review. Scrivener and Word require export to markdown. Proposed advice: Start bounded to avoid invented tags and chaos. Verify quotes against transcripts.
Try it
I want to index my newsletter archive and repurpose one article using Claude Code. My archive is in [replaceable folder path]. My style guide is in [replaceable CLAUDE.md path]. Target output: [replaceable output folder]. First, read @CLAUDE.md and list the top 5 voice rules. Then, scan [replaceable folder path] and propose a structure for indexing 10 recent articles. Do not move files yet. Once I approve the structure, index them. Next, take [replaceable article filename] and generate three repurposed files: newsletter.md (400 words), thread.md (8 posts), and summary.md (120 words). Apply my voice rules strictly. Show me the diffs for each file. I will review and approve each change before saving. If you cannot read a file, ask for the correct path. Do not invent tags. Keep CLAUDE.md under 200 lines. Measure success by: 1) Indexing completed without file loss, 2) Repurposed files match voice rules, 3) Diffs reviewed and approved. If voice rules are unclear, ask for examples from [replaceable voice-examples.md].
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.