I’ve written a lot—hundreds of blog posts, thousands Tweets, and a book about the fundamental uncertainty of all knowledge. Nevertheless, I struggle to write.

My brain is not naturally inclined towards language. I didn’t start speaking until I was 3. I’m more comfortable writing code than prose. And when I write for publication, I spend long hours editing to make my thoughts comprehensible.

Or at least, that’s how it was until about a year ago. Then, I started collaborating with Claude on most of my writing projects. At first I was just as slow as ever, but I eventually learned some useful patterns. Now, working together, we write about three times faster and hit a higher standard of quality than I did alone.

This post documents those patterns so that you might copy them to improve your writing.

Co-Writer, not Ghost Writer

Like I said in my post about writing code with Claude, I think of it as a collaborator and partner, not like a magic box that does things for me. When I sit down to write, it’s like I’m working with the most patient editor who’s ever existed. Claude is always willing to keep iterating on the same text until we get it right and is always available to work with me at any time of day. Some things Claude is especially good at include:

• Fact checking

• Generating variations on existing sentences

• Analyzing structural issues and suggesting alternatives

• Brainstorming

• Researching and synthesizing information

• Writing first drafts that roughly match my style

The workflow looks like this: I bring ideas, goals, and sometimes pre-existing writing. Then, I iterate with Claude in tight loops to organize, refine, and research. We keep going until the text meets or exceeds my standards, and when it seems like we’re done, I do a final readthrough. If everything still sounds good, then we’re ready for publication.

That’s the workflow at a high level, but as always, the devil is in the details, so let’s look at exactly how Claude and I write together, including some example prompts used to produce my recent essay about The Sort.

Subscribe now

The Workflow

1. Start with Brainstorming

Instead of starting with a first draft, like I would have done before working with Claude, I start with brainstorming. I’m not very good at brainstorming, so I get Claude to do the babble, and I do the prune.

Example:

Prompt: i’m writing an essay for my blog about fallout of the sort, which i’m calling living in the shadow of the sort. i want to spend a bit of time exploring some of the consequences of the sort, why it exists, and how we might get out of it (and if that’s even a good thing). don’t try to write the post for me right now, but give an outline of topics you think would make sense to include. right now we’re brainstorming, so err on coming up with too many ideas to include, and then we’ll use that to narrow down to figure out what makes sense as an outline later. the the idea of the sort comes from a patrick mckenzie tweet. here’s the whole thing so you have context: [copy/paste of the twitter thread]

Claude generated a couple dozen potential angles and over a hundred individual points. We then went back and forth to narrow these down and find the right essay structure. This took maybe 30 minutes, but it meant when we started actually drafting, we had a fleshed-out plan.

2. Draft in My Voice

Since I’ve written so much, I have a defined writing style. Whether or not my style is good is up for debate, but it definitely exists. When I draft a post, I instruct Claude to read several things I’ve written. It also helps that my posts are in the training data and it already knows who I am.

Example:

Prompt: okay, great, i think this is a good start. now, you have some samples of my writing and you can find more online. try to convert this into an essay. i think we should target something in the 3k word range, though we can go longer if that seems necessary to explore some particular idea

The output was generally in the direction of my style, but regressed in the direction of Claude’s voice. I needed to do extensive editing to produce something I was happy to publish, but for a first pass, the draft was a huge timesaver.

Note: At this point I’ll put the draft in a Google doc and share it with Claude. This makes it easier when I need it to read the whole thing after I’ve made edits.

3. Structural Review

Next, I read the draft and work with Claude on sections that feel off. I’ll do some of the rewriting myself, but then come back when I notice that a section isn’t doing quite what it should. As is probably obvious, it’s a big time saver to get the structure of a piece right before trying to do stylistic edits on paragraphs that might never appear in the final version.

Example:

Prompt: hey the japan section feels underdeveloped. you set it up as an example of resisting the sort, but then say it’s a combination of factors and move on. i actually think the japan example is an important hinge in the essay. let’s figure out how to make it that.

This prompt saved me from polishing a section that was fundamentally failing to do what it needed to. By noticing early and working to fix it, we saved an hour or more of writing time.

4. Research

I write non-fiction, and that means doing research. I do plenty of this myself, but I need Claude to know about it, too, so I often do my research with Claude, asking it to do web searches, give me links, and synthesize what it finds.

Research can happen at any time. I’ll frequently find in the middle of a piece that I need to know more about something. If I think it’ll be quick, I’ll do the research in the chat I already have going. If it seems likely to be complex, I’ll start a new one, and when done, I’ll ask Claude to give me a summary to paste back into the original chat to hand over context.

Note that asking Claude to do research is importantly different from relying on what it thinks it knows! It’s often critical to include the words “do a web search” in a prompt to avoid hallucinations.

Example:

Prompt: do we have the details about japan right? do a web search to make sure the details about the lost decades are correct and that we aren’t missing anything important that would make this section wrong

This came back with some helpful details that confirmed we mostly had it right in the text but had a few minor details wrong that needed fixing.

5. Iterative Refinements

This is where I see big productivity gains. When I write solo, I’ll spend an hour struggling to fix a paragraph that feels wrong. With Claude, I flag what’s wrong and get several alternatives in seconds. None of these are usually the final version, but they help me make steady progress towards a version of the paragraph I like.

Example:

Me: somethings not working with this paragraph. i think it might sound off because the example is too on the nose. how might we fix it [paste paragraph]

Claude: [provides 4 initial options]

Me: okay, i ran with one of those options. here’s the new version but not quite there yet. suggestions? [paste paragraph]

Claude: [provides more variations]

Me: great, this is getting closer. i like the first two. here’s the new version i wrote, wdyt? [paste paragraph]

…

Eventually we got to something I was happy with. The trick is to never struggle alone. If I feel stuck and don’t know what to do, I talk to Claude about it to generate new ideas that get me unblocked.

6. Fact Checking

I regularly ask Claude to review drafts specifically for errors. This includes both factual errors, like when I get a claim wrong, and consistency issues, like when two parts of the text contradict each other. I ask it to fact check both before and after edits, since sometimes what seems like a simple stylistic change introduces language that makes the literal interpretation of words wrong.

Example:

Prompt: fact check this [paste paragraph]

I then iterate on the text until it passes the fact check. Sometimes this becomes quite involved if I’ve misunderstood something.

For example, while working on book revisions, I once had to spend several days with Claude to better understand the Löbian obstacle. In the end, I’m still not sure I understood all the nuances, but thanks to Claude, I’m pretty sure I didn’t say anything obviously wrong in the text, which is much better than I would have been able to do on my own. Perhaps a patient friend with a deep understanding of Löb would have been better, but I doubt they would have wanted to wake up and write with me at 6 am!

Taste and Thinking

This workflow works well for me. But, similar to what I said in the programming essay, it only works well because I bring good judgement and taste to the writing process. I wouldn’t be able to write the way I do with Claude if I didn’t know what good looks like.

I also need to be able to think clearly. Claude can easily become a mirror for my confusion, and not in a way that helps me become less confused. I regularly have to push back against Claude being too agreeable. It’s not a sycophant, but I can easily bias it in my direction, and sometimes I have to specifically ask it to be critical or to play devil’s advocate.

The good news is that writing is a path to clarifying thinking. As Paul Graham said, “If you’re thinking without writing, you only think you’re thinking.” The catch is, you can’t outsource your writing to Claude and expect to get better at thinking. So while Claude is a powerful tool for clearing writer’s block and providing detailed feedback, you must ultimately do some writing yourself to think and write well.