Voices

A style profile says how far a text is from a corpus; it cannot make the text sound like the corpus's author. A voice can. It is an object of the organisation, assigned per occasion — chat, customers, publications — to an agent, a department or the organisation as a whole, and chosen at the moment of writing by whom the agent writes to (below).

A voice carries:

• a card — 250 to 400 words on how the author opens, carries an argument, brings in facts, builds sentences, and what they never do;
• passages — five to eight paragraphs that show the hand; they go into the prompt;
• a chat tone — address, tone and emoji in the team chat;
• when it is measured from texts: a profile the style gate checks every outgoing text against, and a contrast against AI text if a reference was uploaded.

Which voice applies: per occasion and per reader

An agent does not carry one voice for everything. A voice is chosen per occasion:

• Chat — the team chat, and the work an agent does for a request from a conversation;
• Customers — mails, tickets and replies in target systems; every other task;
• Publications — blog posts, docs, offers; a task says so with a line occasion: publications (a HEARTBEAT.md aufgabe: can carry it too).

For each occasion, the first of three places that names a voice wins, at the moment the agent writes:

1. the department of the person the agent answers — set in the org chart's edit mode under Audience and voices;
2. the agent — its settings, Voices, one slot per occasion; an empty slot says what applies instead and from where;
3. the organisation — Administration → Default voices.

Customers have no department, so for them only the agent's and the organisation's voices count. In a group chat with people from more than one department, no single department's voice fits everybody, so the agent's own applies. When nothing is named, the agent writes as it did before: with the TONE.md in its config.

How to speak with us. A department can add one line — at most 400 characters, for example "figures first, no ticket numbers" — that goes with everything an agent writes to somebody of that department, whichever voice applies. In a group, the lines of every department involved come along together (at most four).

Who gets what. The voices list shows a table: every department, and everybody without one, against the three occasions, with the voice in effect and where it comes from; pick an agent to count its slots too. A voice's own page lists who names it for what, and its preview can be written to somebody of a department. Each chat answer shows the voice it spoke in ("voice: X · for Sales"), and each run records the voice and the reason (agent×customers, department:Sales×chat, none×chat).

Setting a slot writes no config version: the voice acts as it stands when the agent writes, and the recording says which one it was. The style gate, covey/style_check and covey/style_apply measure against the voice of the outward occasion — customers, or publications when the task says so.

After the upgrade (migration 0122), an agent's one voice sits in its Customers and Publications slots, and Chat is empty. The team chat then uses the organisation's chat default if one is set, and no voice otherwise; the chat tone (du/Sie, emoji) still comes from the customers voice until a chat voice is named. Nothing about how an agent talks to colleagues changes by itself.

Building one

Voices in the library has a New voice button that walks through six steps; an existing voice shows the same steps as sections of its page. A rail on top marks each step as done, to do, optional or not yet possible, and the line under it names what blocks the next one.

1. Purpose — blog, support mail, chat, offers and proposals, or other. It goes into every prompt that writes in the voice and picks the kind of sample the preview writes.
2. Source — one of three:
   • From texts. Upload what somebody wrote (files or pasted text). While the texts go in, the page says what they lack: how many more texts of 150 words or more are needed for reliable bands (four in all), which texts are too short to carry a band, which address their reader differently from the rest ("these texts look like a different register"), which are in another language, and whose sentences run far longer or shorter than the others'. Build measures them; the card is one model call.
   • From a description. Say how you write, as you would to a new colleague — how a text begins, how facts come in, how long sentences are, what you never write, du or Sie. One model call writes the card, the passages and a suggested chat tone. The card's claims come from your description rather than from quoted passages, and the page marks the voice described, not measured.
   • From the chat. Ask an agent in the team chat to draft a voice. It interviews you in a few questions or takes texts you give it, and files a draft with covey/voice_draft. The draft appears on the voices page, named after the agent.
3. Material — the texts built, or the description written.
4. Chat tone — optional; a described voice brings a suggestion, which applies once saved.
5. Preview — optional; one short sample on a topic you pick, as a paragraph, a mail or a chat message. Not stored. Once a released card has a new draft beside it, either can be heard.
6. Release — read the card, correct it, or refine it by instruction ("less formal, more concrete numbers"): the revision is a draft shown beside the text before. Nothing of it acts until it is released; a released card is not overwritten by a later build.

Described or measured, and the style gate

A voice built from a description has no profile. Nothing was measured, so the style gate does not check against it: it records that it skipped and lets the text pass. The voice still acts where it matters most — in the prompt, through its card and passages. Add texts to it and build, and it becomes measured: the passages come from the texts, the gate applies, and the released card stays.

An agent drafting a voice

covey/voice_draft {"name": "…", "purpose": "support_mail", "description": "…"} — or "texts": [{"name": "…", "text": "…"}] instead of a description — creates the voice as a draft. It has its own guard-rail subject, covey:voice_draft, so an organisation can gate or forbid it, and every draft is in the recording. The agent cannot release, build or assign a voice: those stay with the people who manage voices.

Who may do what

Every role reads voices. Creating, changing, building, describing, previewing, refining and releasing are for org_admin and agent_owner — the preview changes nothing, but it costs a model call. The same two roles set the voice slots of agents, departments and the organisation, and a department's line; everybody else sees them locked.