There’s something about AI writing that stands out. It’s not bad, exactly. It doesn’t drift into cliche; it’s usually concise, if perhaps a little over eager. But it can feel rather flat. And its consistent use of particular words and phrases, never mind its fascination with em dashes, make it easy to spot. And that, in turn, makes people stop reading.

“So what?”, you might wonder. Far too much AI writing is unnecessary; the office-work equivalent of leaving a child unsupervised with a can of silly string. In both cases you’re guaranteed a mess.

I think that’s a missed opportunity. As with everything I’ve written so far about AI, my excitement isn’t so much that it helps me do things I couldn’t do before, but that it helps me do those things faster, and often better. Sometimes that leads to more output — work that wouldn’t have otherwise gotten done — sometimes it leads to better results. Often it’s both.

And this has been equally true for my writing. Now, I love writing. I enjoy crafting an argument; developing a narrative; picking the right word and putting it in the right phrase. I would never claim to be any kind of serious author, but I think I communicate clearly and with (I hope) a degree of elegance. I really do care about the words I use, and how I use them. I was nervous about collaborating with AI in my written work, about losing my voice to the machine.

My early experiments only confirmed my fears. I could get results, but they no longer sounded like me. Worse, caught up in the moment of creation, I didn’t always notice how much the tone of voice had changed until after I’d sent the email, published the document, or written the LinkedIn post.

So I kept experimenting. I’m still learning, but I’ve now got two approaches that I think work well.

The Scaffolder

I write frequently as part of my day job. Most of that writing — emails, Slack messages in particular — I continue to take care of myself. The work is interactive, AI won’t accelerate it, and I’m not trying to deliver an important message. But some of the writing is important: creating documentation that will be read by tens or even hundreds of people, or templates and frameworks that have to stand up to scrutiny. And here, I’ve found a helpful partner with AI.

I use Claude Code (which, despite its name, is very good at producing all sorts of written material), and start by describing what I’m intending to do. I give it as much context up front as I can, and then I tell it four important things.

1. Don’t produce any documentation without my explicit instruction
2. Start with bullet points to outline a structure
3. Ask me questions where my instructions are unclear, or where you need to know more to make a good decision
4. Let’s work through each bullet in turn. I’ll re-write it in my words, and then you can offer suggestions for improvements.

Items 1 and 2 tee things up for the collaboration: I will usually have some kind of structure in mind, but the AI has seen thousands of documents like the ones I write, and it can usually give me a sensible start point for me to change.

Item 3 is arguably unnecessary, but my experience is that it will tend to make assumptions if you don’t prod it to check. Having it ask also forces you to think harder about what you’re trying to do, and keeps you working collaboratively as part of the process.

Item 4 is the most important for keeping your voice: you are taking the lead on the words used.

In practice I find that I often use a mix of my words and Claude’s suggestions, but by interactively stepping through the document I can keep the structure and tone aligned with how I write unaided. Claude can also jump in and offer suggestions when I can’t find the words, or want suggestions for different ways of saying something, but I retain control over the output.

The Editor

The second approach is lighter weight; less an accelerator than a quality control gate. I use this one when the tone of voice really matters — when it’s genuinely important that the result sounds like me. I’ll use the Scaffolder when I’m putting together a template for my team to use, for example, as the output is corporate and professional, not personal. I’ll use the Editor to help me write content that is clearly going to be associated with me. Important, large-group emails, say, or this blog.

With the Editor I still write the way I used to: I structure the content myself, I write every word myself, and I keep going until I’m happy with my draft. I might turn to AI to get ideas for how to convey a specific point, but through a one-off conversation where I take its ideas and manually translate them into my own words.

When I’m done, I ask Claude to review what I’ve written and provide feedback. The exact set up varies depending on what I’m writing. Here’s one of the standing instructions that Claude has for this blog:

Review. Claude reviews for spelling, grammar, and length. Suggests edits and flags anything that reads off. Also checks that the post includes at least 2-3 actionable takeaways (“so what’s”) — things the reader can try, think about, or apply. These can be explicit (“try this yourself”) or woven naturally into the narrative. Short posts may be exempt, but substantial posts should always leave the reader with something to do. Takeaways should be audience-aligned: relevant to a general business audience, not just technical hobbyists. Where the post topic allows, at least one takeaway should connect to a business or professional context (e.g. using AI to fill capability gaps on a team, justifying work that previously failed a cost/benefit test). Steve approves or adjusts.

This approach has the advantage of keeping my tone of voice while also giving me some accountability, and a second pair of (virtual) eyes. Sloppy writing is highlighted, weak arguments challenged, aimless paragraphs are cut. As with the Scaffolder, I can still ask Claude to make suggestions, but I will almost always rewrite them to make them my own.

Write away, right away

Together, these two approaches have made me a lot more comfortable with using AI as a writing partner. With the Scaffolder I get the benefits of acceleration and a thought partner; with the Editor I get held accountable to my writing goals; with both I get solid feedback on my prose, and an effective copy editor too.

They can also combine to help me get started faster — Scaffolder helping me break through the writer’s block of a blank page — and get finished stronger, with Editor pushing me to hone my message and stick to my brief.

I’m still exploring how to get the most out of AI in this scenario, but I’m much less worried about losing my voice to the machine. What approaches have you tried? And do you think I’m right to be comfortable — can you tell that this post has been through the Editor, for example, or does it sound like me?