What it is not

A vision that only lists ambitions cannot be argued with. These are the refusals, and the places where the whole idea might turn out to be wrong.

Refused on purpose

A drag-and-drop editor. Récit assumes you write. A visual editor is a legitimate product and a different one: its source of truth is a document model, not a file you can diff. A browser preview and a studio are planned — an editor that owns the content is not.

A hosted service. No account, no backend, no project stored anywhere. The command line and the packages run on a machine. Whatever keeps the content is already in place: a repository.

A dependency on a model provider. The scenario tooling produces prompts, schemas and context for an agent; it never calls one. Prompts are data, image generation happens outside, and a generated asset records where it came from. Nothing in Récit stops working because a key expired.

A second way of doing anything. No alternative syntax to please a habit, no second bundler, no configuration that can be written two ways. That costs flexibility and buys the ability to change one thing in one place.

Publishing by a machine. Agents may write, convert and validate. A release is approved by a person.

Probably never

Real-time collaborative editing. It needs a document model with conflict resolution, which contradicts a plain file as the source. Two people edit the same file the way they edit code.

Arbitrary layout. A format decides how a scene is placed. Pixel-level control belongs to a design tool; what Récit offers is themes and layouts you can extend, not a canvas.

Importing other tools' formats. Reading .pptx, .key or .docx means inheriting their model, which is the thing being replaced. A conversion that loses structure silently is worse than no conversion.

Where this could be wrong

Three honest risks, which the project does not get to argue away:

  1. One source may not survive six renderings. A deck wants six words on a slide; a report wants a paragraph; a film wants a spoken sentence. If the same text cannot be all three, the promise degrades into one good format and five mediocre ones. The bridges are where that shows, and they are tested first for exactly that reason.
  2. The twin may be a tax nobody wants to pay. Keeping Markdown and JSON in agreement costs a schema, a converter and a round-trip test per format. If nobody ever uses the JSON, that work was spent on a hypothesis about machines reading content.
  3. The scope may be the point of failure. Nine formats is a lot of surface for one model. The mitigation is that a format is a package that adds a renderer and nothing else — if adding one ever requires touching the model, that is the signal the set is too wide.

What would change our mind

If the bridges turn out to produce output people rewrite by hand anyway, the premise is wrong and the honest answer is to be an excellent Markdown deck tool and drop the rest. That outcome is acceptable. Claiming nine formats while shipping one would not be.