I designed a 0→1 AI-native presentation platform that combines template creation, document-driven generation, and contextual AI editing without asking users to surrender authorship.
Automation feels fast until users have to reconstruct what changed, why it changed, and how to undo it.
One-shot generation created a draft but also created verification work. The system needed to move between suggestion, preview, application, and recovery without breaking the user's mental model of the deck.
A user may want help with one sentence, one slide, the narrative structure, or the entire deck.
Changing an argument can affect titles, supporting content, order, and visual hierarchy elsewhere.
People need to inspect recommendations, understand rationale, edit results, and recover earlier states.
The model delays commitment. AI first scopes the request, generates a plan, exposes the proposed change, then applies it only at the control level the user selected.
The entry prototype separates template-led, document-led, and AI-led creation so the system knows what context it can safely infer.
Before creating the deck, the system previews the planned outline and lets users revise assumptions while changes are still cheap.
Inside the editor, suggestions are scoped to a selection and presented as reviewable changes rather than invisible replacement.
The system asks about audience, goal, and scope when a request could create broad structural change.
Users can regenerate one affected region instead of restarting the entire deck or losing manual edits.
Undo restores the prior state and retains the rejected rationale so the next attempt can avoid the same direction.
The prototype defined what every AI edit needs to carry: target, scope, rationale, dependencies, user decision, and recovery state. This gave product and engineering a stable interaction contract across multiple AI actions.