Sounding Right Isn't the Same as Being Right

By Mariane McLucas | modulemakers.com

What context files can do, and the one thing they can't

There's a post circulating on X that teaches you how to duplicate your brain in Claude. Ruben Hassid calls it exactly that. A hundred questions across seven categories, voice-to-text dictation, three markdown files loaded into Claude Cowork. The goal is to make AI output sound just like you.

It will work. And it will leave you with a problem you didn't know you were creating.

What the Post Gets Right

Generic AI output sounds like nobody. It hedges, it qualifies, and uses words like "delve," then it produces paragraphs that could belong to anyone in any industry anywhere. Context files fix that. Plain language instructions fix that. Teaching AI your rhythm, your vocabulary, and what you find annoying fixes that.

The post is right about all of that.

The Gap It Leaves Open

There is a difference between sounding like you and being reliable for you. That gap is where things can go wrong.

An AI trained to sound exactly like you will also make things up using your voice. Using your tone. Bringing your confidence right along with it. It will cite a statute that does not exist in the cadence you use when you are absolutely certain of something. It will give wrong information in your word choices, your paragraph structure, your sign-off.

The mimicry does not stop at the things AI gets right. It carries straight through to the things AI gets wrong.

And familiarity is exactly what makes errors harder to catch, not easier. The output feels like yours. It reads like yours. Your brain registers it as yours. So you move on.

The Honest Thing About Voice-Matched Output

Good AI output that sounds like you usually sounds like you because you touched it. You changed the punctuation. You cut a sentence. You used didn't instead of did not. You added the specific observation that only you would make because you were there and the AI was not.

That is not automation. That is collaboration. The output sounds like you because you revised it, not because a language model cloned you.

What Context Files Actually Need to Contain

Most context file approaches teach AI your voice. Tone preferences, vocabulary, what you find annoying, and how you structure paragraphs.

What they almost never contain is reliability instruction. What to do when uncertain. How to flag confidence. What it is not permitted to present as fact. When to stop and ask rather than proceed and guess.

Those are different categories of instruction and they produce different outcomes. A context file that covers the first and ignores the second gives you AI that gets things wrong in a way that sounds exactly like you. Your tone. Your confidence.

And when it makes something up, that made-up fact carries your name.

The Part No Context File Solves

No matter how detailed your setup, no context file fixes drift. It does not address hallucination. It does not curb the quiet confidence with which AI produces false information, because that confidence is a feature of how these systems work, not a gap in your prompting.

The answer is not a better brain duplicate. The answer is a verification layer running underneath everything those files do.

That layer is less satisfying to build. It doesn't produce impressive before-and-after screenshots. It mostly produces AI that admits when it does not know something, which is unsexy and completely essential.

What This Post Is Not Saying

This is not an argument against the approach. If you want AI that sounds like you, Ruben's system works. Build it.

Just don't stop there.

Making AI sound like you is the easy part. Making AI be right is the work.

—

Mariane McLucas is a senior instructional designer and solo consultant at Module Makers. She is developing a training program for professionals who want to use AI reliably in their real professional workflows. Details at modulemakers.com.