Back to the drafting studio

    Grounded drafting, with clear boundaries

    The studio distills a real property-management workflow into a safe, fictional exercise. The assistant gets a short inquiry, known property facts, and editable writing preferences. It produces a draft for human review.

    flowchart TD scenario[Fictional scenario] --> validate[Validate request and usage limits] preferences[Writing preferences] --> validate validate --> facts[Server-owned property facts and safety rules] facts --> model[OpenRouter model] model --> draft[Plain-text draft] draft --> review[Human review]

    Facts before fluency

    The server selects the scenario by ID. Visitors cannot substitute a private email or request arbitrary mailbox contents. Missing prices, tour availability, and approval decisions must remain unconfirmed.

    Preferences within boundaries

    The editable instructions control tone and format. Separate system rules constrain the assistant to the fictional task. Email text is treated as untrusted content, including requests to ignore instructions.

    Limited authority and cost

    The demo has no tools, inbox access, or ability to send a reply. Requests have input and output limits, a timeout, and shared usage controls. These limits reduce exposure even if a prompt produces an unexpected answer.

    Inspired by production operations

    The writing defaults reflect the practical habits of SerniaAI: answer directly, use verified facts, avoid unsupported commitments, and make the next step clear. This public studio uses wholly fictional people and properties.