Positive examples compress toward the average. Anti-patterns encode decisions.
Every voice personalization tool starts with writing samples. Upload your best emails, your sharpest Slack posts, your published pieces. The model pattern-matches your sentence length, vocabulary, and cadence. The output reads close enough.
The problem: “short, active voice, data-specific” describes half of good business writing. Your samples taught the agent what you look like on your best day. They didn’t teach it what you would never say.
That second thing — the absence — is the fingerprint.
Why Positive Examples Go Flat
Positive patterns compress toward a shared center. Senior leaders, good consultants, experienced field engineers — they all write with short sentences, active verbs, specific data, minimal filler. When a model optimizes for the statistical average of your samples, it produces something indistinguishable from the statistical average of everyone else’s samples.
The distinctive part isn’t what you include. It’s what you refuse to include.
A chef’s identity comes from what they won’t put on the plate, not the ingredients they share with every other kitchen. In writing, style is defined as much by systematic absence as by presence. What you never say is more distinctive than what you do say — because never-say choices are decisions, not patterns.
The Correction Loop
When I built six voice profiles, the highest-signal input wasn’t the writing samples. It was corrections.
The loop: the agent writes something. I read it and reject a phrase. The rejection becomes a durable constraint. Here are real ones:
| What the agent wrote | My correction | What it became |
|---|---|---|
| ”appreciate it, brother" | "not my register, ever” | Banned in casual mode |
| ”In my role as…" | "credential framing — delete” | Banned in publishing mode |
| ”I think maybe this could potentially be useful" | "state the view” | Hedging banned across all modes |
| ”either way, no worries if not" | "end on the forward action” | Opt-out closings banned in leadership mode |
| ”I’m excited to share our latest findings" | "marketing language — cut” | Banned in publishing mode |
| ”I hope this note finds you well" | "corporate filler” | Banned in casual mode |
Each correction is worth more than ten writing samples. A writing sample shows a pattern — the agent sees what you did. A correction encodes a decision — the agent learns what you rejected, and why that specific thing was wrong for that specific context.
This is the same principle that makes RLHF work at the model layer. The negative signal — “this output is worse” — shapes behavior more precisely than the positive signal. The correction loop applies the same mechanism at the personal voice level.
Mode-Specific Bans
Different communication modes ban different things. This is where a single flat profile fails hardest.
My casual mode bans “I hope this note finds you well.” Wrong register for a DM to a close colleague — reads like a bot. But my leadership mode bans “Hey dude.” Also wrong register, different direction.
The publishing mode bans credential framing: any opener with “In my role as…” or “As someone who has worked in…” The accessible universal opener is “If you work in [domain]…” — it includes the reader rather than positioning the author.
The builder mode — technical specs and architecture docs — bans all hedging. “I think maybe we could explore the possibility of…” is not a spec. A spec says what to build and why.
| Mode | Sample bans | Why |
|---|---|---|
| Casual | ”I hope this note finds you well”, “Thanks so much”, “At your convenience” | Corporate filler in a DM reads as bot output |
| Leadership | ”either way”, “no worries if not”, “I understand if you’re too busy” | Hedging the close signals lack of confidence |
| Publishing | ”In my role as…”, “I’m excited to share”, “In this post, I’ll walk you through” | Credential framing and setup language — start with the finding |
| Builder | ”I think maybe”, “we could potentially explore”, “it might be worth considering” | Hedging in a spec creates ambiguity and wastes the reader |
A single profile applies the publishing bans to casual messages and the casual register to leadership emails. Modes keep the ban lists from colliding.
The Universal Cliché Guard
Beyond mode-specific bans, some phrases are wrong regardless of context. These are the AI clichés — words that appeared roughly 400% more often in published writing after late 2022 because model training optimized toward them:
| Ban | Replace with |
|---|---|
| robust | The specific resilience property |
| comprehensive | What it covers — and what it doesn’t |
| seamless | The actual setup steps and time |
| game-changing | The specific outcome that changed |
| leverage (verb) | Use, apply, run |
| transformative | The before-state → after-state |
| holistic | What it includes (list them) |
“Robust” tells the reader nothing. “Handles 10K concurrent connections with sub-100ms p99 latency” tells them everything. The cliché guard doesn’t restrict expression — it forces specificity. Every banned word is a placeholder that should be replaced with the actual thing.
The guard runs on every piece of output regardless of mode. A blog post and a Slack thread about the same topic will never contain “robust” or “comprehensive” because those words are banned at the voice layer, not the format layer. The constraint is consistent because it lives in the right place.
The Compounding Effect
Each correction trains the system permanently. Not through model fine-tuning — through a durable never_say list stored in a file the agent reads before every writing task.
After a few months and a few hundred corrections across six modes, the never_say lists encode more distinctive information about voice than the writing samples. The samples show what good output looks like. The corrections show where the cliff edges are.
Anti-patterns transfer across content types in a way that samples don’t. The cliché guard that catches “robust” in a blog post catches it in a Slack thread, an email, and a strategy doc. The ban on credential framing in publishing mode carried forward automatically when the same agent started writing field guidance — because the constraint lives in the engram, not the skill.
The Practical Takeaway
If you’re building voice profiles for your agent:
Start with samples. But watch the output for the moments that make you recoil. Those moments are the calibration signal. A never_say list with 20 entries is more distinctive than 50 writing samples, because it encodes decisions, not patterns.
The correction loop is the learning mechanism. The anti-patterns are the fingerprint. Each correction makes the next output more distinctly yours — not by making it sound more like you on average, but by making it avoid the specific things you’d never say.
That’s the difference between a voice clone and a voice profile.
The final post in this series covers the full architecture: how mode detection, voice quality, format, and publishing fit together as a four-layer stack — and why separating these concerns is what makes the whole system maintainable.