What the AI sees

Sent with every build message

• Your message, with each link replaced by the text of the page it points at (fetched and extracted on this Mac first), and each attached file’s text (PDFs, text, HTML, Markdown) or picture (fit within 1,568 px, as JPEG).
• The shape: every node type with its properties and kinds, which property is its title, and every relationship type with the pairs it connects, in the names you see.
• The earlier turns of the conversation that came to something: your messages, Claude’s summaries, and what you did with each draft.
• While a draft is open: the draft as it stands, so “add the authors too” revises it rather than starting over.
• With Existing nodes on: the names and types of your nodes (all of them up to a limit; for a larger graph, only the ones your message seems to mention, found on this Mac by name and by meaning, up to a cap), so drafted relationships can point at them by id.

Never sent: node contents beyond those names, your notes, your history, anything from other projects.

What it’s asked to return

A structured draft: nodes (each with a ref unique in the draft, a type, and properties as name/value pairs, every value a string: numbers as digits, links as full URLs, booleans as true/false, an enum as one of its options), relationships (type, from, to, by ref or by the id of an existing node, with properties), changes (action remove/update, a description, a read-only Cypher query returning the records, properties to set), and a summary for you.

It’s told to use type, property and relationship names exactly as listed, to invent none, to give every new node a title, to connect to existing nodes rather than create duplicates, and that relationships are directed and allowed only between the listed pairs. Anything that still doesn’t fit is reported under “didn’t fit your shape”.

Thinking

The model’s reasoning streams into the panel as it works, under the drafting step, so you can watch it weigh a sentence. It’s kept with the conversation.

Through the command line

Each call is one run of Claude Code or Codex in a non-interactive mode with a structured-output request. The model id and effort go with it; the tool reports which model actually answered, and the meta line shows it. If the tool is signed out or too old for the model, the turn fails with the fix offered.