Voice calibration deep-dive

Every draft prompt includes a voice context block built from three inputs: writing samples, tone slider values, and banned phrases. Understanding how they combine helps you tune them to match the output you want.

The voice context block

It looks roughly like this at the top of every draft prompt:

VOICE CONTEXT:
Tone axes: formal 22, professional 35, serious 18, curated 12
Distinctive: uses concrete numbers; self-deprecating asides;
  sentence fragments for emphasis; comfortable with profanity
Writing samples (3):
  ---
  [first 300 chars of sample 1]
  ---
  [first 300 chars of sample 2]
  ...
Banned phrases (never use, even in paraphrase):
  unlock, leverage, game-changer, synergy, excited to share, thrilled,
  rocket science, circle back, low-hanging fruit, ...

How each input shapes output

• Writing samples anchor the model to specific phrasings and sentence rhythms. This is the strongest signal. Models match style from examples better than from abstract descriptions.
• Tone axes nudge when samples are thin or conflicting. With 3 samples and clear axes, output is consistent. With 1 sample and conflicting axes, output drifts.
• Banned phrases are hard constraints. We also add them to every follow-up prompt and re-check output. In practice, clean enough to rely on for "gross SaaS speak."

Per-app voice

Each app has its own voice. Useful when you run a B2B tool (more formal) and a personal newsletter (rawer) in the same account. Settings → Customize → Voice within that app workspace.

Refreshing voice

If your writing style evolves, add more samples. The analysis runs each time you add one and blends it with prior samples — not a replacement. You can also force a re-analysis (Re-analyze voice) which discards prior and starts fresh.

What voice calibration does NOT do

• It doesn't fine-tune the model. We use prompt context only.
• It doesn't share your samples with provider partners.
• It doesn't generate identical outputs — samples are inspiration, not copy sources.