Field notes · FN-013 · Per-language tone profiles — what they actually look like, with examples
Drawing G2-FN-013 · Field journal · entryTone & voice

Per-language tone profiles — what they actually look like, with examples

By Andrew Maryasov · Grow2.ai2026-03-21 · 8 min
Entry FN-013
Filed 2026-03-21
Read 8 min
Category Tone & voice
Languages — EN
Entry №
FN-013
Field log

Iryna Skoryk, Grow2.ai's tone editor, runs a Day 5–6 workshop on every pilot. The output is a 4-to-7-page tone document, one per language the agent will speak, written with the client's actual front-desk lead. This note shows what one of those documents looks like and the three first-draft mistakes that show up on every project.

Why per-language, not just per-brand

A Polish guest at Atlas Rentals expects formal address ("państwo", conditional verbs); a Ukrainian guest expects warm directness (active voice, second-person singular "ти" if under 35, "Ви" otherwise); an Italian guest expects context-rich greetings, even at 03:00 reporting a broken kettle. A single "brand voice" document cannot encode all three without becoming uselessly abstract.

The structure of the document

  1. One paragraph: who are we talking to? Demographic specifics, urgency profile, what they're typically holding when they read the message (phone in airport, laptop at desk, dental chair).
  2. Three example exchanges: ideal greeting, ideal apology, ideal escalation refusal. Written by the front-desk lead first, then edited by Iryna for AI-friendliness.
  3. Five "never say" phrases. Real examples from drafts that got vetoed. ("Як справи?" from the agent in UA — patients found it weird.)
  4. Address rules: formal/informal, naming conventions, when to use the agent's name vs. the brand name.
  5. Time-of-day modifiers: what the 02:14 reply sounds like vs. the 14:02 reply.

The three things that always go wrong on the first draft

Grow2.ai has now written tone documents for thirty-one pilots. The same three issues appear in the first draft of nearly every one:

  1. Em-dash overuse. The default Claude/GPT tone leans on em-dashes — they sound thoughtful in English but read as "trying too hard" in Polish, German and Ukrainian. We strip them out by hand in the prompt.
  2. Apologies that don't apologise. "I understand your frustration, however…" is a textbook AI apology. It contains the word "however", which signals to the reader that what follows is a deflection. The replacement: "You're right that this should've worked. Here's what I'm doing about it now."
  3. Wrong register on first contact. Atlas's guest is on holiday; the agent should not open with formal "Dear Sir or Madam". Clinica Via's patient is anxious; the agent should not open casually. The first line of the system prompt encodes register explicitly.

Have a problem this note describes? Bring it to a call.

Field notes are written for the version of Grow2.ai that will run into the same problem in eight months. If one of them describes your situation, that's usually a good sign we should talk.

▸ Commission a pilot
Per-language tone profiles — what they actually look like, with examples · Grow2.ai field notes · Grow2.ai