Repository navigation
Text to graph design
Give an AI a piece of text and get back a graph of the things, kinds and links it names. It comes as a proposal you accept or reject. Designed on 18 October 2026, against version 0.5.0, at Roger's request. Not built yet.
The requirements are section 47 of
requirements/editor-requirements.gql,
slice 10, in workstream X1. They are built in Wave 19 of
requirements/waves.json, after the
Obsidian community release (R2). Roger's rule is that what he adds now goes after R2.
| Door | Where | How it sends |
|---|---|---|
| In the editor | A command on selected text or a whole note | The AI send rules of §19: off until chosen, the model for proposing changes, within its ceiling, seen before sent, recorded |
| As a skill | A Claude skill whose SKILL.md also loads in Copilot Cowork | The conversation's own model; the library proposes the block, nothing is written before approval |
Both take the prompt and the answer's shape from the library released on GitHub, so the same text gives the same kind of proposal everywhere. The model's answer is untrusted. The graph format library reads and checks it before anything is shown, and a bad answer is refused with the reason.
- Each new thing, kind and link, marked as an addition (§17).
- Things already in the note's graph, matched by name or alias and marked, never added twice.
- For each item, the words of the text it came from, so it can be checked against the source.
- §17 proposals and §19 AI rules, already built for the AI send dialog.
- §46, the Copilot Cowork skill, which ships the same SKILL.md as a
.skillwith each release. - §45
board_ai_decompose, which does the same for a whiteboard; text to graph is its counterpart for prose.
Using it
The project
Building it
Designs