Measured #1 · July 2026
We measured the “voice guidelines” method.
It’s better than you think. Until turn three.
There’s a workflow every AI-productivity channel teaches right now: paste a few of your writing samples into ChatGPT or Claude, ask it to write “voice guidelines” describing your style, save those guidelines in a Project, and let every future draft inherit your voice. It’s free, it takes ten minutes, and it’s the main alternative to tools like ours. So we tested it the way we test everything: with numbers, published win or lose.
Spoiler for fairness: the method partly won. Then it lost in a specific, repeatable, architectural way.
Short one-shot tasks: a straight tie
Three short tasks (status email, LinkedIn post, declining a meeting), fresh conversation each time. DIY guidelines averaged a 90 style match against the persona’s baseline. Our engine also averaged 90. For a four-sentence email in a fresh chat, a good model following fresh guidelines is genuinely enough.
Long-form: both methods degraded, ours included
On a ~720 word newsletter, both methods slid to 75: each let uncontracted phrasing creep in as length grew, and the DIY tail sprouted a classic AI tricolon (“we’re learning. we’re listening. and we’re building this with you”) in a persona that never writes as “we”. Long-form voice is an open problem for everyone, and we publish our own 75 rather than hide it.
Multi-turn: the guidelines decay, and this is the real finding
Real people don’t open a fresh chat per email; they keep working in one conversation. Three sequential tasks in a single thread: the DIY method scored 100, then 80, then 78 as its own previous outputs piled into the context and diluted the saved guidelines. Measured generation, which re-reads the samples and re-scores every single request, went 94, 100, 100. The drift isn’t a prompt-quality problem. It’s what happens to any static instruction as context accumulates, and it’s why measurement has to happen per request, not once at setup.
The numbers
| Test | DIY guidelines | Measured generation |
|---|---|---|
| Short tasks (mean of 3) | 90 | 90 |
| ~720-word newsletter | 75 | 75 |
| Conversation turn 1 | 100 | 94 |
| Conversation turn 2 | 80 | 100 |
| Conversation turn 3 | 78 | 100 |
Method and disclosures
The method under test is the workflow taught in popular tutorials such as Eddy Ballesteros’s guide and “Make AI Agents Write EXACTLY Like You”: samples in, AI-written voice guidelines out, guidelines saved as standing instructions. We followed it faithfully, including letting the model write its own guidelines.
Fairness controls. Both arms used the same base model, the same four writing samples, and the same tasks. The only variable is the method: static saved guidelines versus per-request measurement with a scored revision loop.
Our biases, stated. The ruler is our own deterministic scorer, the same one documented in how the style match score works and runnable by anyone. We build the competing method’s opponent, so treat this as a vendor’s experiment with open math, not a neutral paper. The test persona is a constructed, disclosed fixture (a consistent casual founder register), not a real person’s private writing. Sample size is one persona, one model, three turns: directional, not definitive. We’ll gladly publish reruns that contradict us.
Run the comparison on your own voice
Paste one AI draft and one real email into the free demo. No signup. If the DIY method serves you better, use it, and now you know where it breaks.
Try the free demoMore measurement: what statistically gives AI writing away