Propose a Map from Canon
3 min read
Sync gives you the structure your team already wrote down in folders and requirements. It cannot give you the parts of the product nobody filed a folder for.
For that, Atlas reads Canon — your specs and PRDs — and drafts a map from them. Nothing it produces reaches the map until you approve it.
Two Modes, One Action
Section titled “Two Modes, One Action”The Discover action in the rail does two different jobs, and its label tells you which one it is about to do.
| Map state | Label | What it does |
|---|---|---|
| Empty | Read Canon and propose the product’s areas, screens and routes | Drafts a whole map from the corpus |
| Populated | Find gaps — read Canon again for what the map is missing | Diffs Canon against your curated map and offers only the difference |
Use Discover once, early. Use Find gaps whenever Canon grows.
Generate a Proposal
Section titled “Generate a Proposal”-
Make sure the project has documents in Canon that reached Ready. With an empty Canon there is nothing to read.
-
In the Atlas action rail, select the Discover action.
-
The proposal generates in the background. The rail shows it working.
-
When it is ready, select Review what Canon proposed.
How Canon Is Read
Section titled “How Canon Is Read”Atlas does not use the same retrieval that grounds test case generation. It enumerates every Canon document by section heading instead.
The reason is measured. On a 226-document project, retrieval selected 12 documents — 5% of the corpus — and produced a map of five areas, two of which were a troubleshooting guide and a best-practices guide, while the document whose headings are the product’s navigation never made the cut. Headings are short enough that the same budget buys the whole corpus.
Retrieval is still the right instrument for detail on a single node. It is the wrong one for the skeleton.
Review the Proposal
Section titled “Review the Proposal”The review panel exists to make accepting a deliberate act.
- Nothing is accepted by default. A proposal is a suggestion from a model that produces a different tree on every run.
- A title already on your map defaults to Merge, not Add — so you cannot create a second “Checkout” by clicking through.
- Implied nodes are flagged. A node is marked implied when its title does not actually appear in a cited heading — the model grouped it itself. Look harder at those.
- Unsupported nodes are shown, not hidden. What the model wanted to say but could not cite is still a question worth asking a human. It just is not evidence.
Each proposed node carries the Canon citations it was drawn from, so you can check a proposed area against the document that implied it.
Decisions
Section titled “Decisions”Set one decision per proposed node:
| Decision | Effect |
|---|---|
| Add | Create this node on the map |
| Merge | Fold it into the existing node with the same title |
| Skip | Discard it |
Decisions are saved as you make them, so a long review survives closing the panel.
Apply or Discard
Section titled “Apply or Discard”Apply writes the accepted nodes, routes and journeys to the map. Accepted nodes are marked as AI-accepted, which exempts them from future syncs — sync will never rewrite or remove them.
Discard throws the whole proposal away. The map is untouched.
When a Proposal Fails
Section titled “When a Proposal Fails”If Canon is too thin to map confidently, the panel says so rather than inventing structure. The fix is on the Canon side: add the documents that describe the product’s shape — see Canon best practices.