Claude Opus 5.5 Drafts, This Writer Cuts 8 of 10 Words

Picture a new hire on day one. They give you their first draft, and you go through it with a red pen until there’s more red than black. On day two the draft is a bit better. By week six you’re barely touching it, because they’ve picked up your taste just from watching what you cross out.

One LinkedIn creator used exactly that picture to describe how they now write a newsletter for nearly a million readers. I couldn’t stop thinking about it after I read the post. The workflow is smart, but the mindset behind it is what really got me.

The author’s newsletter has about 996,000 subscribers. For the latest issue, Claude Opus 5.5 wrote 2,303 words. The author kept only 438 of them word for word. The author didn’t even count that themselves. They asked Claude to do it.

That means the author cuts about 8 out of every 10 words. Most people would call that a failed draft. The author treats it as a lesson.

The twist: every deletion is feedback

This is the part that changed how I think about AI writing. Every word the author removes teaches Claude something about what they want next time. The edits aren’t wasted work. They’re the training.

All of it happens on GitHub. If that word makes you nervous, relax. The author says they made the setup non-technical, and they’ve never written a line of code. Better still, they never open GitHub. They only talk to Claude, and Claude handles everything behind the scenes.

Think of GitHub here as a memory bank. It keeps every draft, every edit, and every version. Over time that history becomes a detailed record of the writer’s taste, and Claude can look back at it whenever it writes something new.

So who’s actually writing the newsletter?

This is the obvious question. If an AI wrote 2,303 words, is it still the author’s newsletter? The author says yes, without hesitation, and gives 4 reasons:

1. They pick every idea. The author is the spark: a conversation they had, a thought, something they read.
2. The opinions are theirs. Claude doesn’t decide what the newsletter believes.
3. They rewrite most of it. About 80%. The author says you can check this yourself on Pangram, a tool that detects AI-written text.
4. Writing is thinking. You can’t outsource your thinking.

Then comes the line I keep coming back to:

“Claude can’t judge my writing. It can only learn from my edits.”

That one sentence reframes the whole AI writing debate. Claude isn’t the judge of what’s good. The human is. Claude is a fast, tireless apprentice that improves only because someone keeps marking up its work.

The red pen loop

Back to the new hire. The author says their Claude now works a lot like one. It learns their taste from their red pen, and GitHub keeps everything stored so no lesson gets lost between sessions.

The author also repeats one point so nobody misses it: they never open GitHub. Only Claude. The technical layer exists, but the writer never has to deal with it directly.

The author calls this the most transparent newsletter they’ve ever written, and by far the most technical setup they’ve ever shared. Their walkthrough is titled “How I write a newsletter to 996,007 people, with Claude Code & GitHub.” They promise you can build the same thing even if you’ve never written code. By the end of their guide, your Claude will tell you where it needs you.

Why this matters beyond newsletters

Most people use AI like a vending machine. You type a prompt, take what comes out, and either ship it or give up on it. Whatever you learn from the result disappears when you close the chat.

The author’s approach flips that around. The gap between what the AI wrote and what you published becomes the most valuable data you have. It applies to almost any writing you do over and over:

• Sales emails: your edits show Claude which claims feel too pushy for your brand.
• LinkedIn posts: your cuts show which hooks sound like you and which sound like a template.
• Client reports: your rewrites teach the tone your clients actually respond to.
• Internal docs: your trims teach Claude how short your team really wants things.

How to start your own red pen loop

The author’s full walkthrough covers the exact setup. In the meantime, here’s how I’d apply the core idea starting today:

1. Write down your own idea before you open Claude. One line is enough. It keeps you as the spark, just like the author.
2. Let Claude write a full first draft from that idea and your opinions.
3. Edit hard. Don’t be polite. Cut anything that doesn’t sound like you.
4. Ask Claude to save both versions, the original draft and your final edit, in one place like a GitHub repository, so the history builds up.
5. Ask Claude to study the difference with a prompt like: “Compare your draft with my final version. List the patterns in what I cut and what I kept, and update your writing notes for next time.”
6. Track how much of the draft you keep, like the author’s 438 out of 2,303. If that number rises over the weeks, the loop is working.

A few tips to get more from it:

• Don’t fix things in your head and retype them somewhere else. The edit has to land in the saved version, or Claude can’t learn from it.
• Keep the ideas and opinions yours. The author is clear that this is what keeps the work authentic.
• Expect a low keep rate at first. Cutting 80% is part of the process, not a sign it’s failing.

The point isn’t getting AI to write for you. It’s building a system where every edit you make makes the next draft better, while the thinking stays yours.

I love how honest this post is. Plenty of people quietly use AI and hope nobody notices. This author published the numbers, the reasoning, and the setup. If you want to see how the system is built, read the full LinkedIn post.