Are three texts enough to build a reference voice?
Three texts can support a first measurement. They cannot summarise a person. The reference depends mainly on what those texts share, and what they are not expected to represent.
By ShakuPublished Updated
A minimum reference is not a complete identity
Three coherent texts can already reveal stable habits and support comparison with a fourth. That is enough for a useful first hypothesis, but not to lock an entire voice into a permanent portrait.
A newsletter, an internal memo and a video script may come from the same person while using different registers. Mixing them without naming those contexts blurs the reference instead of enriching it.
Choose texts that have a reason to be compared
The best corpus is not the one containing your most impressive work. It represents the voice you want to recover in one specific situation.
- choose texts you still recognise as your own;
- avoid work substantially rewritten by someone else;
- keep the register comparable: newsletter with newsletter, speech with speech;
- prefer texts long enough to reveal a rhythm;
- add new examples over time instead of searching for a perfect corpus at the start.
The reference becomes more precise over time
Kaneme treats the reference as a living instrument. The first texts set a direction; later validated texts strengthen stable traits and reveal sub-registers.
The useful question is therefore not “do I have enough texts to define who I am?” but “do I have enough comparable texts to make a first useful decision?”
Read next
Measurement · 6 min
Same brief, two voices: what the measurement actually sees
Two texts can answer the same brief without carrying the same voice. Here are the signals Kaneme compares, and those it refuses to invent.
Portability · 6 min
Carry your voice into ChatGPT, Claude and your agents
A portable voice is more than a prompt copied everywhere: it combines a common reference, an operating contract and measurement after generation.
Measurement · 8 min
Revise AI-assisted writing without disguising it: a method and six prompts
A guide to rewriting an AI draft in an identifiable voice without trying to evade a detector.