I’m writing a book. I started working on it way too long ago. Years ago. The first draft was almost complete when I got an opportunity to lead a new project, and as I often do, I thought I could handle both on top of my day job. As I often realize, I couldn’t. Something had to give, and so the almost-ready first draft remained just that.
Until about a year ago, when I decided this was one loose end I had to tie up. I was proud of what I had managed to achieve in that draft, so why not finish it and make it publication-worthy? And so, I opened my archived project and read the manuscript.
It wasn’t good. First drafts rarely are — that’s why they are drafts. But this one caught me by surprise. I didn’t expect it to be perfect, but I also didn’t expect to stop at every paragraph and wonder what I was thinking. I knew the ideas were good, but the execution wasn’t. For some strange reason, this made me even more motivated. Editing my book was going to be a project in itself, and I love projects. So, I had a plan for that summer.
Summer passed. Autumn was gone. Around November I kinda realized this was going to be as difficult as writing the first draft. Maybe even more. A year ago, I didn’t imagine I would still be working on my book in May 2026. But I am. And to my surprise, I find it both thought-provoking and highly enjoyable.
With AI helping so many people achieve a lot more much faster, you’d expect a project like this to take a fraction of the time. Instead, AI slowed me down. I wanted it to. And in the process, I learned (and I’m still learning) things that go way beyond this project. I am relearning the craft of writing.
Act 1: The Spellchecker
English is not my native language, yet almost everything I write, I write in English. Somehow, it always felt more natural. Which is not to say that my writing is flawless. Far from it. From silly typos to grammar mishaps, from failing to choose the right word to getting the tense wrong, my writing is never publishable as is. And that’s why, from the first blog post I ever published, the spellchecker was my safety net. Theoretically.
Back in the early 2000s, the built-in spellcheckers in word processors were the most accessible solution in the wilderness of writing tools, and all they could do was mark misspelled words. Some word processors went as far as introducing grammar checks, but they were so basic that a simple second read could have caught most, if not all, of the errors they flagged. And that’s assuming that these were actual errors. The few standalone tools in this arena generally didn’t do much more.
Then came Grammarly and enabled me to see the oceans of issues I missed in my writing. It was no longer the basic in-app spellchecker with its facade of a grammar validator. Grammarly was nuanced enough to give me the sense that I was improving my text for real. Was I improving my writing skills? I honestly can’t say. While Grammarly provided a much more elaborate safety net, it was still just a safety net.
Safety nets are important, and you can learn from errors. Yet mistakes are not the only source of learning, and not everything worth learning is right or wrong. As good as Grammarly was, it focused on what I did wrong, not on what I did right. Even when the style layer was introduced, it was always binary: better do this than that. With tools like Grammarly, there was no way to learn from the gray area — where there’s an actual dilemma and things are not black or white. The gray area where the most meaningful learning emerges.
Your Daily Dose of Creativity
Download the CreativityMEDS iPhone App
Act 2: Chatting with Claude
When ChatGPT was launched in November 2022, everyone was amazed by its capability to generate textual content. “Nobody will ever have to write anything anymore!” was a typical reaction, at least in tech blogs and podcasts. I wasn’t impressed. The technology was jaw-dropping, but the use case didn’t appeal to me. I wanted to write, and I didn’t want anyone or anything to do my writing for me.
Claude was launched a few months later, and I experimented with both not long after. But it took me almost two years to realize I could use Claude to rethink how I write: I now had someone to discuss my text with. Discuss what? Well, mostly my spelling and grammar.
So in parallel to using Grammarly for flagging the more obvious (and at this point, quite repetitive) issues, I started to ask Claude to highlight where I could improve my writing. It gladly pointed out spelling errors but was also able to provide me with a long list of grammar and style issues. Some of them were plain errors, but others were more nuanced. Instead of accepting or rejecting a change, I could ask a follow-up question. I could ask Claude to explain the root of the issue. Claude always offered a revised version for a sentence, but in most cases, I didn’t just use it as-is. Instead, I came up with my own revision and was able to get immediate feedback on how well it resolved the issue.
From the very start, I knew I had to insist that Claude not just play along with whatever I wrote or proposed. I wanted it to push back, but I also wanted it to be open to “being convinced” and supportive when I actually did something well. And while no LLM is perfect at providing unbiased, objective feedback (and frankly, most humans aren’t either), having feedback I could engage with was itself a significant improvement. My goal was not to revise my text as fast as possible. My goal was to learn.
But the experience was still limited. First, it covered only spelling, grammar, and style. It was also restricted by the size of the text. Claude would happily find twenty issues in a short blog post, but when I gave it a complete chapter from my book… it came up with more or less the same number. Its attention span was finite.
Yet the biggest problem was structural: the insights sat in a separate list, and I couldn’t see them in context or keep track of multiple discussions on different issues. Grammarly’s insights, limited as they were, sat overlaid on my original text. That alone made for a better experience than scrolling through Claude’s separate list and monolithic chat.
When I got to try Claude Code, I knew what my next step would be.
Act 3: My Editorial Team
Yes, I love projects, and one project at a time will never do. So when I saw an opportunity to experiment with Claude Code, I wanted to check just how far it could take me. I wanted a better way to engage with the feedback, but more importantly, I was looking for more than just grammar and style analysis. I wanted the opportunity to review and improve complete chapters of my book and consider aspects I hadn’t until that point. I decided to take a short break from my writing project to build my editorial dream team.
My typical editorial flow runs through multiple lenses, each focused on a different dimension of the text. It starts with the question: Have I achieved the declared goal for my predefined audience? The second editor maps the text’s DNA: it highlights where the form supports the message and where it’s working against my goal. One editor steps into the role of a knowledgeable, skeptical reader — stress-testing the logic of my arguments and probing for gaps or unsupported claims. The fourth editor analyzes my writing craft and the impact it has on my readers. Next, my team highlights potential style issues and flags grammar errors. The rhythm of the text is analyzed to identify monotonous parts and repeating words. Finally, the fact-checker validates claims I make and references I use.
Here’s an insight one of my editors flagged as something to rethink about the paragraph you’ve just read:
Editor:
The editorial-team framing is the essay’s most ambitious workflow claim, but p24 — the only evidence paragraph — lists the editorial lenses without showing one in action.
A single concrete example — one editor’s insight, one back-and-forth — would close the gap between claim and demonstration. As written, this is the essay’s weakest evidence-to-ambition ratio.Me:
This is a bit tricky to demonstrate without the reader having the context of the analyzed text.
What do you think about quoting this insight on p24 as the reader just read it and can judge the editorial comment in its context?Editor:
That’s a smart instinct — and structurally it could work well, but I’d push back gently on the placement. Quoting an editorial insight on p24 itself would interrupt the lens-cataloguing rhythm […]
A cleaner structural fix: insert a new paragraph between the lens list and what follows — one that zooms in on a single lens in action with a short quoted insight on the very text the reader just finished reading.
The risk is tonal: if the quoted insight is too flattering it reads as self-congratulation; if it’s critical, you’d need to show what you did with it. A back-and-forth fragment (insight → your response → revision) would be ideal but might be too long. A single insight plus one sentence on how it shaped a revision is probably the right size.
And so, I’ve added these fragments of our dialogue.
How good are these insights compared to a human editor’s? That’s a real question. At minimum, they are more effective than those of any other non-human editor I’ve come across. The point is that even if the insights are merely good, or just adequate, this process does something real: it gives me feedback and allows me to discuss it.
Each of the editorial modes produces a thorough analysis, sometimes with dozens of insights. None of them change the text automatically. None of them just propose an alternative. My editorial team always explains what I did well and what could be improved. At times, my editors offer two points of view without recommending which I should adopt. When I read these insights, I sometimes know immediately what I need to revise. In many other cases, the insights my co-editors generate are just the starting point for a discussion.
My co-editors use what I write to help me become a better writer.
From Generative AI to Editorial AI
We tend to think about AI through the lens of productivity: a tool that can help us produce more for less. It’s an acceleration tool. What once took months can now take days, and what required a team can now be done by one or two people. I can’t deny that. A year ago, I couldn’t have imagined building something like the CO.EDITOR app myself. Claude Code made it possible, and it did so in a fraction of the time it would’ve taken me to develop this workflow alone.
But that’s not the only way to think of AI. AI can also slow us down and add friction. Not the friction that wears you down — the kind of friction that makes meaningful learning possible. Friction that invites you to think and not automatically accept. Shortcuts and productivity hacks are great, but some things require time. Mastering a craft is one of them.
My editorial workflow slows me down, as any editorial process should. It is not optimized for speed — it is optimized for quality. It is not designed to get more things done, but rather to help me get things done better. To help me get better.
There are many areas where I do use AI for speed. Writing code is one of them. Before I started working with Claude Code, I struggled with the question: Is this something I am willing to delegate? I started as a software developer, and for years this was part of my identity. Ultimately, the answer in my case was yes! Why? Because in retrospect, writing code isn’t a skill I aspire to master. That’s not a direction I would’ve invested in as part of my personal journey. Developing software is not what defines me anymore.
Writing, on the other hand, is. Being able to capture my ideas in writing — to organize, phrase, and communicate them so others can understand what’s on my mind and build on it — is a skill I would never give up or delegate. Never. And that makes the slower path more valuable. That makes the investment of time and effort worthwhile. And if Claude can help me take this slower path by providing insights, engaging in thoughtful discussions, and offering professional guidance, there’s no reason not to take advantage of it.
We try to push productivity to its limit with AI. We rarely stop to think about the value of a slower path. The alternative starts with mapping the skills you consider essential. That’s not always a trivial task, but it is critical. The next step is to hold on to them. Don’t delegate these skills. Continue to develop them. No matter how much time and effort are required. And if you can use AI to support you in this journey, by all means, use it to create friction. Use it to slow down.
CO.EDITOR (Beta) is available for free. Using it requires your own Claude API key, for which you will be billed by Anthropic per usage.
If you just want to get an impression of what it can do, you can explore the built-in examples without an API key.