How I used Claude Skills to make AI sound like me
TL;DR:
This post is about the tension I keep hitting when I use AI for writing: wanting it to make my work better without it quietly replacing the way I actually think and talk.
It's about what I learned when I went looking for a technical solution, and why building something that sounds like you is harder than it should be. If you've ever finished an AI-assisted piece of writing and thought, "This is good, but it's not really me," this one's for you.
There's a tension I keep coming back to when I use AI for writing.
On one hand, it feels respectful to the reader to put your work through AI. To make sure it's sharp, well-structured, and actually worth their time. On the other hand, it feels disrespectful to hand over something I can't genuinely stand behind and call my own words. I find myself caught between those two things more often than I'd like to admit.
A little while ago, I spent almost a full day writing a LinkedIn post about starting Brightside. I'm not a super confident writer, so I brought Claude in to help me sharpen it. What came back was well-written, but it wasn't me anymore. Every time I tried to pull it back, it got further from where I started. Eventually, I gave up on the whole thing.
When I came back to it a couple of days later, I'd watched a video about someone using Skills in Claude, and I wondered if that was the answer. Not using AI to write for me, but finding a way to inject my voice into the system so that when it edited my work, it wasn't quietly removing all my stylistic choices along the way. I think that's one of those areas where cognitive surrender can really take over. I wasn't confident, so I was treating AI like the expert. But you always have to be in the driving seat.
So what are Claude Skills, exactly?
They're roughly equivalent to Gems in Gemini or GPTs in ChatGPT, but with one key difference: they're stackable. Instead of jumping between different tools to get different things done, you can draw on multiple skills within the same conversation. A brand voice skill and a content creation skill, for example, working together in the same piece of work. That's what makes them interesting.
I found a skill called ‘Humanizer’ on a website called Skillhub, where anyone can browse, rate, and download skills that other people have built. I knew enough from digital marketing to be cautious about downloading things from the internet and adding them to my system, so before I did anything with it, I copied the text into Claude and asked it to tell me what the skill was actually doing and whether there was anything in there I should be worried about.
Claude's verdict was that it was well put together, but it probably wasn't going to do what I wanted. The issue was a section that instructed the skill to "add soul," which was really about adding a generic human voice. Not my voice. Just a voice. And that tension was exactly what Claude would have been wrestling with when trying to execute what I was asking it to do.
That conversation taught me a lot about what actually goes into a skill. They all need a specific name and a description that tells Claude exactly when to use it and what it does. The instructions themselves are detailed and very specific, with lots of examples of what good looks like, what bad looks like, and how to troubleshoot when things go wrong. The more specific you are, the better the output. That part felt familiar, honestly. It's not that different from briefing a person.
There's also a Skill Creator skill, built by Anthropic, that walks you through building your own from scratch in plain language. You don't need to be technical. What you do need is a clear sense of what you want the skill to do and what good output actually looks like for you specifically.
What I built instead
In the end, Claude and I didn't create a full skill. We created a document called Sarah's Voice, a detailed set of instructions for how Claude should write when it's writing as me. Claude's suggestion was to practise with it first inside a project, test the output, refine it, and then develop it into a proper skill once I was confident it was doing the right thing.
Finding source material that was genuinely, unambiguously mine was harder than I expected. Anything recent had already been through AI at some point. I ended up going back to a pet-sitting manual I wrote for Humphrey's carers (RIP beautiful dog), some old cover letters, a few emails. It was a bit of a time capsule, but it worked.
I'm still in the testing phase. But what I've learned is that the goal was never to hand my voice over to a machine. It was to make sure that when I work with AI, it's a genuine partnership. My ideas, my words, my stylistic choices, with AI acting as a skilled editor that knows when to improve something and when to leave it alone.
That's a harder thing to build than I expected. But it's the right thing to build.
If this resonates and you're trying to figure out how AI fits into your own workflow without losing what makes your communication distinctly yours, that's something we think about a lot at Brightside. Get in touch, and we can talk through it.
Published by Brightside Collab | Written by Sarah Croney
