When the Voice Guide Becomes the Problem
I wrote a voice guide to help AI match my writing style. It worked too well—the AI learned the example phrases, not the principles behind them. Here's how I fixed it, and what it taught me about the difference between describing a voice and understanding one.
Nino Chavez
Product Architect at commerce.com
I wrote a voice guide to help AI match my writing style. It worked. Too well.
The guide had examples. “Ask me again in six months” as a way to end provisionally. “I used to think X, now I think Y” to show evolution. “Two different modes. Same instinct.” for punchy fragments.
The AI learned the examples. Not the principles behind them—the actual phrases. And suddenly every draft ended with “ask me again in six months.” Every post had an “I used to think” moment. The fragments felt manufactured.
The guide had become a template. And templates, applied consistently, create tells.
What I Was Actually Trying to Teach
The voice guide captures how I write. But it captured the what more than the why.
When I write “ask me again in six months,” I’m signaling something: this isn’t a final answer. My thinking is in motion. The reader should hold these ideas loosely.
That spirit can be expressed infinite ways. Ending mid-thought. Posing a question I haven’t answered. Simply stopping without a tidy conclusion. The phrase was one instance of the pattern, not the pattern itself.
Same with “I used to think X, now I think Y.” The spirit: show that my perspective has shifted. Let readers see the evolution, not just the current position. But the explicit before/after structure isn’t required. Sometimes the shift is implied. Sometimes the tension between old and new thinking lives in the argument itself, never named directly.
The Session That Broke It
I ran a voice check on a draft about Vercel’s Agent Skills announcement. The draft hit all the marks—question-first opening, self-interrogation section, provisional ending.
Too many marks.
“Ask me again in six months” at the end. “Here’s where my thinking has been evolving” in the middle. The structure was correct. The voice was hollow.
I’d trained the AI to pattern-match against my guide’s examples. The guide had become the problem.
The Fix
Three changes to how the guide works:
1. Spirit descriptions before examples
Every pattern now starts with why it exists, not what it looks like. The evolution pattern isn’t about saying “I used to think X.” It’s about showing that your perspective has moved—through whatever language achieves that.
2. A “retired phrases” list
Some phrases get used so often they become tells. “Ask me again in six months” is one. “Here’s where I’ve landed—for now” is another. These were authentic once. Through repetition, they’ve gone stale. The guide now explicitly retires them.
3. A freshness tracker
New patterns get overused too. The guide now tracks what’s appeared in recent posts. If I used “here’s where my head is” last week, I shouldn’t use it again for a month. The tracker creates forced variety—not by restricting what I can say, but by noticing what I’ve already said.
The Rewrite
Original ending of the Agent Skills draft:
That’s the frame I’m holding, anyway. Ask me again in six months.
Revised:
But that graveyard keeps pulling me back. All those ideas that died not because they were bad, but because the execution gap was too wide. If we can narrow that—even partially—we unlock something.
Not replacement. Expansion.
Same spirit. The door stays open. The thinking feels unfinished. But the words are different.
The original ending was a formula. The revised ending earns its openness through the argument itself.
What This Is Really About
The deeper issue isn’t phrase repetition. It’s the difference between describing a voice and understanding it.
A voice guide that lists examples teaches imitation. The AI (or a ghostwriter, or future-me on a tired day) can hit every checkbox and still produce something that doesn’t sound right. Because the examples were never the point. They were evidence of something underneath.
The revised guide tries to teach that underneath. What is provisional language for? What does showing evolution accomplish? Why do fragments work when they work?
When you understand the function, you can find your own forms.
The Meta Layer
There’s something recursive about using AI to help document how to work with AI. The session that produced these insights was itself an agentic workflow:
- Voice check surfaced the problem (pattern-matching to phrases)
- I diagnosed the root cause (examples without principles)
- The AI helped draft guide revisions (spirit-first descriptions)
- We rewrote the draft together (applying the new approach)
- The comparison revealed what changed (and what still needs work)
The AI wasn’t autonomous. But it wasn’t just executing commands either. It was a collaborator in the meta-work of improving how we collaborate.
What I’m Still Figuring Out
The freshness tracker is an experiment. I don’t know if forced variety produces better writing or just different writing. There’s a version of this where the tracker becomes its own constraint—avoiding patterns that would have been the right choice because they appeared too recently.
And there’s a question about whether voice guides should exist at all. Maybe the right answer is fewer explicit rules, more examples to absorb implicitly. Maybe codifying voice is itself the problem.
I don’t have that resolved. But the draft reads better now. The ending doesn’t feel like a template. The evolution shows through the argument instead of announcing itself.
That’s something.
This is the fourth entry in the Agentic Workflows in Practice series. The first three covered video compression, Claude Cowork, and document generation. This one covers the recursive case: using the workflow to improve the workflow.