Three models, one deck: how Decklyn writes, designs, and narrates

A presentation is three things at once: words, pictures, and a voice. For a long time software could help with maybe one of them. Now three kinds of models, language, image, and voice, are each good enough that you can hand the whole job to them and get something you would actually stand up and present. The interesting part is making them agree.

Step one: write the deck

First a language model turns your topic and audience into structure: a title, an argument that builds slide by slide, tight copy on each slide, and the words a narrator would say out loud. We keep this strict. The model returns validated JSON, and we hold a plain rule: draft, do not invent. It writes from what you gave it and marks figures as things for you to check, it does not fabricate a statistic and dress it up as fact.

Step two: design each slide

Then an image model generates a visual for every slide, from a short description the writer produced, in one consistent house style so the deck looks designed by a single hand. Bright, clean, on your brand color. If a visual does not land, you regenerate just that one, and if the model is unavailable the slide falls back to a tasteful gradient rather than breaking.

Step three: give it a voice

Finally a voice model reads each slide's narration in a natural, professional voice. That is what turns a deck into a talk: you can present it live, or share it as a narrated deck someone watches on their own time. The three steps run in order, and you can edit any of them.

The product is not one model. It is the seam where three of them meet.

We keep the language layer model-agnostic on purpose, so we can move to a better model without rewriting the product. The image and voice layers are the same idea. The point is the result: a topic in, a designed and narrated deck out.