Skip to content

What runs your label ​

A model answers you. Everything around it — the job descriptions, the knowledge, the counting — is ours, and you can read it.

A language model, in one paragraph ​

A language model is a program that predicts the next piece of text, given the text before it. Google's own introduction puts it plainly: it "estimates the probability of a token or sequence of tokens occurring within a longer sequence of tokens" (Introduction to Large Language Models). A token is a word or a piece of one. Trained on an enormous amount of text, that prediction gets good enough to answer questions, follow instructions and write a draft. Claude, GPT and Gemini are language models sold by Anthropic, OpenAI and Google (Intro to Claude).

Two things follow. The model has no memory of you between conversations unless we send it something. And it is good at writing, not at counting or knowing: it will write a convincing deadline that is wrong. Both shape how the label is built.

An agent is a model plus a folder ​

Each department is one folder. The model reads it at the start of every conversation (How it works):

PartWhat it isWho wrote it
The job descriptionthe role, what it owns, what it never does, how it talks to youus
The skillsone task each — "rank the radio stations", "draft the pitch" — as stepsus
The scriptsthe counting, dating and ranking, as code that gives the same answer every runus
The knowledgewhat the department knows about the industry, each rule with its source and a last-checked dateus, from the primary pages
The testssimulated artists hold whole conversations with the department before a real one doesus

The model supplies the language: it reads your message, follows the skill, calls a script when a number is needed, opens a knowledge page when a rule is needed, and writes the reply. It does not invent the rule or the number — see What the agents never do.

Which model ​

You choose a provider once, at setup. You never choose a model: the label picks one per phase, from three tiers.

TierUsed forWith an Anthropic key
quick back-and-forththe interview questions, routingClaude Haiku 4.5
everyday worka department doing its normal jobClaude Sonnet 5
the heavy thinkingthe release plan, the presskit, the pitchClaude Opus 5

Those are the models the label pins today for Anthropic. Anthropic's current line-up also includes Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 (Intro to Claude); a monthly check flags when a pinned model has a newer sibling, and an upgrade goes through the tests first. The other providers have their own three models each — see Providers, keys and cost.

Anthropic is the only provider the departments have been tested on. The instructions were tuned on Claude; a model that follows them loosely gives a worse interview, not an error. The setup says so when you pick another one.

On a Claude subscription the label runs inside Claude Code, and the model is whichever one your Claude Code session uses.

What one conversation sends to the provider ​

Every turn, the label sends the model one package. It contains:

  • the department's job description and the skill it is running;
  • the titles of that department's knowledge pages — a page is sent only when the model opens it for your question;
  • your artist file: your profile, and what this department has remembered about you;
  • the roster, so a department can point you to another one;
  • today's date, and everything said so far in this conversation.

It does not contain the whole catalog (only the rows a search returned), your Google Drive (only a file a department reads for the task), or any other conversation. The package is a few thousand words and is identical from one turn to the next, so the provider caches it instead of re-reading it — which is what keeps a conversation cheap. Where each piece lives, and what never leaves your machine, is on Your data.

Next: Providers, keys and cost.

Sources ​

Beta · Quebec validated · Canada and the US next.