← Web app

The build assistant

The build assistant is an optional hosted chat available on the Assistant tab of a created build. It helps you draft the working fill by reading the template brief, searching your assets, and writing suggested values — using the same draft-validation path as any external agent or API call.

How it works

1. You type a message on the Assistant tab — typically a brief description of the deck you want (topic, audience, key points).
2. The assistant reads the annotated catalog (the field names, descriptions, and agent hints you set during annotation) and the current working fill, then calls a read-only asset search tool to find relevant images, charts, or other media from your library.
3. It replies with a chat message and, when it has enough information, writes a complete updated fill into the build. That fill goes through the same draft validation as any other write — the assistant cannot bypass it.
4. When the assistant needs more information it may ask structured questions, rendered as an inline answer form in the chat thread. Fill in the answers and submit the form to continue.

Honest caveats

• No streaming. Each turn is a synchronous round-trip. Replies appear after the full response is ready.
• Conversations persist per build. Closing the tab and returning later resumes the same thread. Conversations are scoped to one build and are not shared across builds.
• The human presses Build deck. The assistant only edits the working fill; it never starts the build engine. Review the Slides and Values tabs before clicking Build deck.
• Server-side AI is opt-in. Simply not using the Assistant tab sends no data to the assistant’s model provider. The fill builder (Slides and Values tabs) works entirely without it.
• Rate-limited. Turns are rate-limited per user to prevent runaway spend.

The underlying model

The assistant uses Gemini (default gemini-2.5-flash, configurable via ASSISTANT_MODEL). See specs/user-experience/assistant-spec-v1.md for the full technical spec.