Product

How brand voice works in Kiln

What regeneration does, why prose matters, and how proposed updates protect your edits.

What the brand profile is

Kiln builds two layers from your sources. The brand profileon the Brand page is yours to edit — descriptors, audience, do/don't lists. The voice spec behind the scenes holds corpus-grounded evidence: stylometry, verbatim exemplars, and recurring phrases pulled from your writing.

Drafts read both. You see the editorial layer; the model also gets rhythm and real sentences from your corpus.

What regenerate does

Clicking Regenerate from sources re-runs analysis on your ingested chunks. Examples and the voice spec refresh automatically.

If you have already filled in the brand form, Kiln proposes diffs instead of overwriting — you accept or dismiss each field on the Brand page. An empty profile is applied in full on the first run.

Why prose matters

Voice is inferred from prose sources: websites, file uploads, and Hacker News — not from Python or TypeScript files. A repo-only workspace gets hedged voice fields and no exemplars until you add writing.

See the sources guide for what to connect.

When voice feels off

  • Thin corpus — one marketing page is not enough. Add essays or newsletters you have published.
  • Repo only — add a website or upload before expecting a recognizable voice.
  • Stale sources — regenerate after major new writing is ingested.
  • Hand-edits vs proposals — check for pending suggested updates you have not accepted yet.