How polish actually works.

Crestio uses Anthropic Claude. Lighter sessions go to Claude Haiku (fast, cheap, plenty for a 60-word note). Anything that needs more reasoning — lesson plans, longer notes, the in-app assistant — goes to Claude Sonnet.

Anthropic is SOC 2 Type II audited, and under their commercial terms your prompts and outputs are not used to train their models. We send only what's needed: the rough note, a brief description of the student (year level, subject), and the most recent few notes for that student so the new note reads continuously.

The first polish you ever do uses a generic warm-but-precise template — almost identical to the homepage demo. It's good enough but it isn't you.

Every time you send a polish, Crestio quietly takes the version you actually sent (after any edits) and adds it to a small style profile for your account. After about 20 sent polishes, the next polish starts using your phrases, your sentence rhythm, the specific way you describe progress and homework. Most tutors stop editing after that point.

The profile lives only in your account. It isn't shared between tutors and it isn't sent to Anthropic for training.

Polish receives: the rough note you wrote, the student's first name and year/subject, and a short window of recent session notes for that student.

Polish does not receive: the parent's email or phone, internal tutor-only notes (a separate column with separate visibility), payment information, or any other student's data.

The polished output is a draft, not a sent message. You see it in a preview pane with the rough version side-by-side, and you can edit any character before pressing Send. AI output can be wrong; you're the last review.

There's an undo window after send — about 8 seconds. After that, the parent has the message in their inbox.

It will, sometimes. The most common failure modes: misreading a student name as a topic, omitting a homework note that wasn't formatted as a list, or being too warm when the rough notes were straightforwardly factual.

If a polish is genuinely wrong (a quality issue, not just a style preference), email lenin@crestio.ai with the rough note and the polished output. That's how the prompts get tuned.