Your AI Voice Guide Probably Isn't Doing What You Think (Run the Control Test)

Posted on Oct 20 • Originally published at engineering.kenmazaika.com

My agent handed me a draft that opened with a scene from my own life:

On October 19, I wrote down one rule for every overnight OpenClaw job I run: "work on this overnight" is a request, not a scheduler.

I never wrote that sentence. There was no October 19 writing session, no notebook, no rule I sat down and recorded. The agent invented it, because my voice guide told it to open on a concrete, countable instance of my own work.

That's a bug in the file, but it isn't the interesting one. The interesting one is that I had no way to know whether the file was doing anything at all — and I'd been running it for a month.

Twenty-odd rules. A paragraph cap. A seven-question sign-off at the end of every draft. Everything I published went through it, and every time a draft came out better than the last one, I credited the guide.

I never ran the same sample with it and without it. Not once.

Without a control, a guide gets credit for everything: a draft improves and the guide did it, a draft gets worse and the topic was hard. That's not a voice system. It's a mood with a filename.

Hold everything still, change one thing.

The Reddit post is the subject. The essay is the sample. Same model, same brief, same day.

The no-guide version is genuinely good. It opens on the lesson, uses two em dashes, runs sentences from eight words to thirty-five, and invents nothing.

The full guide produced five sentences averaging 32.6 words, zero em dashes, and a fabricated opening. It isn't sloppy — it's compliant. The self-specificity rule demanded a concrete personal moment and the no-fabrication rule forbade inventing one. The first rule won, because a concrete first sentence is checkable and "don't invent a memory" isn't.

The trimmed guide lands on my published shape: mean 16.6 against a corpus mean of 15.2.

The trimmed guide beat no-guide on two things: it opened on the number instead of the lesson, and it stopped sooner. That's the whole delta — about thirty seconds of a reader's attention.

So the defensible claim is narrower than the one I'd been making:

A voice guide is a filter for AI idioms. It is not a transfer function for your voice.

Mine caught the phrases I hate — the "it's worth noting," the forced triads, the recap conclusion. That's a real job. It just isn't the job I thought I was paying for.

What survived is small enough to read in one sitting, which is the point:

Minimal voice guide The five spine rules 1. Open on the concrete thing. The first sentence names the tool, number, file, or result.

No framing, no backstory, no "in today's…". 2. One idea per section.

If a later passage repeats an earlier beat, cut the later one. 3. Receipt, then verdict.

Name the artifact or number, then one line on what it means. A stat never ends a paragraph. 4.

Operator-active first person. "I ran it, it did X." Not "it returns / it packages." 5. Stop when the lesson is done.

No recap, no FAQ, no motivational close. The three voice rules 6. No AI idioms.

No "not X but Y" / "it's not just X, it's Y", no forced triads, no em-dash pileups. An em dash only when it replaces a comma or colon and makes the sentence clearer; never as a rhythm tic. 7.

One owned opinion, stated plainly. No both-sides mush. One inline caveat max, only where a claim needs it.

8. Let sentence length follow the thought. Short punch, then a longer explanation.

Don't perform burstiness, and don't use rhythm as a substitute for a point. The never list Words: leverage · utilize · delve · seamless · robust · holistic · transformative · game-changer · revolutionize · unlock · "in today's…" · "it's important to note" · "at the end of the day" · "dive in" · "in conclusion". Moves: "not X but Y" / "it's not just X, it's Y" · forced triads · em-dash pileups · rhetorical-question filler · recap conclusions.

The three questions 1. Does the first paragraph name the concrete thing? 2.

Does every claim have a receipt or an owned opinion? 3. Would I say this out loud to a peer?

Enter fullscreen mode Exit fullscreen mode Three things got cut, and each one is a lesson:

The kit is a blank template, one Reddit post, one prompt that writes the essay three ways, and the worked example. Clone it, fill the template with the rules you already use, run the prompt, and read the three essays side by side.

If your guide can't beat the no-guide version, its rules haven't earned their place. Now you know that in about twenty minutes instead of a month.

Test the file before you trust the file. A voice guide that has never been run against its own absence is a preference you're calling a system — and the cost is invisible, because it's paid in your cadence.

Originally published at engineering.kenmazaika.com/blog/give-your-agent-your-voice.