Every speaker discovers, sooner or later, that giving a presentation is mostly a question of survival. This cockpit is the closest thing yet to a Babel fish strapped to your slide deck: speak, and titles materialise. Pause, and The Guide whispers the quote, the data point, the counter-argument you almost remembered. One shareable link throws the same talk on every screen in the room. The Guide is definitive. Reality is frequently inaccurate. Mostly harmless.
You don't read from a script — the model still drives every
slide title via tool calls. But if you drop a `.md` that follows
the template (frontmatter + <!-- slide: TITLE -->
markers), it gets attached as context so the model can
anticipate topic shifts and pick titles that match your planned
arc. Don't have one yet? Copy the AI prompt and paste your raw
script into Claude or ChatGPT — it returns a properly-formatted
file you can drop straight into the cockpit.
The cockpit listens to you over WebRTC and asks GPT-realtime to drive slide titles via function calls. A separate long-context agent (gpt-5.4, a million tokens deep) watches the same transcript and surfaces quotes, data points, and counter-arguments into The Guide — never coaching, only substance the speaker can defend. The audience tab is a read-only Stage you can throw on a TV; viewers can also type the four-letter code at /join. Time is an illusion. Slide-prep time, doubly so.