The AI models behind Worsley

Worsley splits the work so each part is done by the right tool, and so no AI ever sets your score.

Jev reads your prompts

Your captured sessions go to Jev, a decision model from TypeSafe AI. Jev doesn't write text. It answers a fixed set of questions about the session, such as what kind of task it is, what each prompt does, and which of the seven components each prompt needed and gave, with a confidence for each answer.

Jev reads your prompts, the names and types of files you attached, and short excerpts of the AI's replies, which help it tell what each prompt is doing. It grades each prompt from that prompt, its files and the earlier turns of the same task, never from the reply to it. When it isn't sure a prompt gave a component, the prompt gets half credit for it. When it can't tell what kind of task a session is, or what a prompt does, Worsley doesn't guess: the session or turn shows as not scored.

A fixed formula scores them

Worsley's own formula turns Jev's answers into your ROIP. Its weights are fixed and the same for everyone, so the same answers always give the same score. See How session scores are calculated.

A language model writes suggestions, when you ask

Insights in words come from Claude, a model from Anthropic, only when you ask for them. Claude words the advice; it never grades. Once written, an insight is saved, so opening it again doesn't generate it again. What it's given depends on what you ask for:

• Feedback on a turn: that prompt, a short excerpt of the reply, the names and types of attached files, and what Jev found.
• Insights on a session: every prompt in the session (shortened), each one's grade and role, whether each reply landed, and your own averages from the last 30 days. Never the replies.
• Tips to Improve: your averages over the last 90 days, your ROIP by kind of task, and the prompts behind your lowest- and highest-scoring turns across your sessions. Never the replies.

Your data

• Your conversations are never used to train generative AI.
• Jev and Claude get your sessions and scores only: no timing, and nothing that identifies you, such as your name or email. Neither ever sees the contents of files you attach.
• TypeSafe processes your sessions only to answer Worsley's questions, keeps them only as long as it needs to provide that service, and doesn't train on them.
• Anthropic doesn't train its models on data sent through its API.

The Privacy Policy has the details.