I’m not a developer. I just published something I built with AI.

Before publishing RawVoice, I had to ask AI to teach me how GitHub worked.

I already had an account. That didn’t mean I knew how to prepare a release, package the files or make sure someone else could download and use what I had built. I needed help with all of it.

I want to start there because it would be easy to share the finished link and leave that part out.

RawVoice 1.0 is now public. It helps ChatGPT and Claude write more like you, using a reusable profile of your writing preferences. I built it with AI’s help, without being a developer.

The story behind it includes quite a few moments of saying, “That’s wrong. Try again.”

I think those moments explain the project better than the download link does.

I know when writing feels wrong to me.

A sentence can be grammatically correct and still sound like something I would never say. An introduction can be polished and give me no reason to keep reading. An email can explain the request while sounding unnecessarily distant.

I don’t need to be a developer to notice that.

When we were preparing RawVoice’s description, the first version was technically informative. It explained the product. But I looked at it and thought: this is corporate and boring.

That mattered. I was building something intended to help people escape generic AI writing, and its own description sounded generic.

I wanted plain words. A recognizable frustration. A reason to care.

“Tired of AI slop?” was much closer to what I meant.

That became a useful test for the whole project: could I explain what I was building in words I would actually use?

I wanted my preferences to survive the next conversation

The idea behind RawVoice is simple.

You tell the AI what sounds like you: the tone you prefer, the phrases you dislike, how you open an article, how you disagree with someone, what changes between a personal story and a professional email.

Those preferences become two files: a detailed voice profile and a short voice card.

You review them, keep the approved versions in a private Project, and use them to guide future writing.

That is the part I wanted to make useful. A writing preference should be something you can inspect and keep. You should be able to see whether the AI understood you, correct it when it didn’t, and carry the file into the next conversation.

RawVoice works through a Project setup in ChatGPT and through an installable skill or Project setup in Claude. The files can move between them, although they don’t sync automatically.

Underneath the packaging, it is a practical way of giving the AI clearer instructions about how you want to communicate.

Building it meant learning how to judge it

I used AI to help develop the instructions, organize the package, prepare the documentation and work through the release process.

My role was to keep checking whether the result matched what I had asked for.

We tested the interview with fictional personas. We checked whether all 15 Quick questions were asked, whether the answers survived in the profile, and whether a new chat could use the saved preferences.

Some of the problems were surprisingly small on the page and surprisingly important in meaning.

“Usually” could become a rule without exceptions. An unknown preference could become a prohibition. Information that hadn’t been supplied could turn into a confident claim that it didn’t exist.

Those differences are easy to overlook when the writing sounds fluent.

They also explain why building with AI still requires attention. I could ask it to generate files or suggest a fix, but I had to decide whether the output was faithful, useful and honest.

Being able to describe what is wrong is a skill. So is recognizing when a proposed improvement has changed the meaning.

I brought that judgment to the process even when I needed help with the technical steps.

The hardest decision was accepting what I could promise

At one point, the testing led me to question whether we could fix the accuracy problem completely.

We kept finding ways a model could introduce something unsupported. We could improve the instructions, preserve distinctions more carefully and test fresh responses. But the user’s prompt would still matter. The model could still make a mistake.

I had to be clear about the promise.

RawVoice helps with your writing voice. It does not remove your responsibility to read the output and check the facts.

That boundary helped me move forward.

I didn’t want to publish something that claimed to make AI infallible. I wanted to publish something useful, explain its limitations and let people decide whether it helped them.

The final ChatGPT test completed the full 15-question interview, produced both profile files and retrieved the saved preferences in a fresh chat. Separate chats produced an email and a fictional LinkedIn post that preserved the supplied facts in those examples.

That gave me evidence of a working process. It didn’t give me permission to promise perfect results.

Using it is much simpler than building it was

You download the version for ChatGPT or Claude and follow the setup guide.

Then you can take the 15-question interview, bring an existing voice profile or share examples of your writing. You review the profile and card, save them in your private Project, and start asking for drafts.

An email. A post. An article. Presentation copy or speaking notes.

You still supply the substance. You still decide what deserves your name.

For me, that is a productive use of AI: give it a defined job, provide the context it needs, and keep your judgment involved. Build a process you can reuse instead of treating every response as a finished answer.

You don’t have to start by building a product. You could start with one task you repeat and one clear description of what a good result would look like.

That is a manageable place to begin.

Publishing it means something to me

I am not launching RawVoice as a developer pretending the process was effortless.

I needed help. I asked basic questions. I rejected outputs. I changed my mind about what the product should promise. And I learned enough along the way to get it into a form other people can try.

I’m proud of that.

The finished project matters to me, but so does the fact that I followed it through. An idea became instructions, files, tests and eventually a public release.

RawVoice 1.0 is available on GitHub. It is free under the MIT license, with setup guides for ChatGPT and Claude.

Try it if you want your AI-assisted writing to sound closer to you. Read what it produces. Correct it. Keep what works.

A short while ago in this process, I was asking how to publish anything on GitHub.

Now I have something there with my name behind it.

I’m still not a developer. But I have built and published something I wanted to exist.