Open-source example
Prism Atlas
A novel-writing app where the manuscript — spine, prose, characters, and their links — is one structured Prism file a writer and an AI agent edit together.
A novel as a structured file
Prism Atlas is an open-source application for writing novels — and the example that shows Prism on the narrative side of the world. Where Prism Light is scenes and graphs and Prism Pixel is dense images, Atlas is the case that looks least like "structured data" at first glance and turns out to fit best: a manuscript is a deeply interconnected project, and Atlas treats it as one.
It is also the most on-message example in the whole Foundation, because Atlas was designed from the start around the exact thing Prism exists to enable: a writer and an AI agent editing one live, structured document together. The original Atlas already ships that experience with a bespoke document model and its own embedded agent server; rebuilding it on Prism is the natural next step — the bespoke pieces become the Foundation's.
What a novel actually is
Under the prose, a novel in Atlas is a graph of small, typed, linked things — exactly the shape of a Prism file:
- The spine —
chapter→outline→beat, each level an ordered expansion of the one above; one altitude per column in the editor. - The prose — a chapter's text as
paragraphs ofsentences, kept beside the spine, not instead of it. - The links — a paragraph realises one or more beats; a sentence depicts events; a word-range links to an entity. These are connections between paths, not prose conventions.
- The world —
entity(character · location · item · event · concept),relation(knows · loves · opposes · locatedAt …), and appearances tying an entity to a beat — a graph over the same prims.
Mapped onto Prism, every one of those becomes a prim with typed properties, and every link a connection. The manuscript grid, the entity map, and the dataflow an agent reasons over are all the same prims, read three ways — the model's central claim, in a domain no one expects it to hold.
What it shows off
- Explicit links, nothing inferred. A paragraph→beat link or an entity mention is a real connection you can query and traverse — "show every sentence that depicts this event," "which beats has this character appeared in" — instead of a full-text guess.
- Branching version history. Atlas keeps a version tree for every editable piece, so nothing is ever lost and authorship (writer vs. agent) is recorded. That is time and identity on the model, made into a first-class writing feature.
- One mutation surface. Every change — typed in the editor or made by the agent — goes through a single entry point, which is precisely the command-bus pattern the Foundation provides.
- Live co-editing with Claude. Atlas exposes its live manuscript to Claude Code so the writer watches the agent read and edit the novel in place. On Prism that is simply Prism Server — the bridge — pointed at the manuscript file.
Why it matters as an example
If a novel — the most human, least tabular artifact imaginable — is cleanly expressible as a typed object model that an agent can safely edit, the positioning generalises: Prism is for complex project data of every kind, not just the obviously-structured kinds. Atlas is the proof at the far end of the range.
Status
The original Atlas is real and in use today, built on its own document model and agent server. Prism Atlas is the planned open-source rebuild on the Foundation — replacing the bespoke document with a Prism file, the custom interface with Prism UI, and the embedded MCP server with Prism Server. It will join Light and Pixel as a complete, readable application you can study end to end — the narrative point of the triangle.