Keeping AI workflows editable

An agent can create a good first draft in seconds. The harder problem begins after a person opens that draft, changes it, closes the editor, and expects the agent to understand what happened.

For OpenGraph Creator, I chose to make the handoff explicit. The agent writes a layered .ogdoc file and a small local session record. Studio edits that same document. When control returns to the agent, it reads durable state instead of guessing from a screenshot or relying on conversation history.

The document is the contract

A final PNG is useful to a browser, but it is a poor collaboration format. Flattening removes names, layers, editable text, and the reasons behind the composition. A structured document keeps those decisions available.

The contract has to stay small. It should describe the canvas, layers, assets, and export intent without trying to encode every possible editor feature. Versioning matters because both the CLI and Studio need to reject structures they do not understand.

Local state should be inspectable

Keeping sessions in the repository-adjacent filesystem has practical benefits:

• a person can inspect and remove them;
• an agent can resume without a provider API key;
• recovery does not depend on a chat transcript;
• tests can exercise the whole lifecycle with fixtures.

This approach also creates responsibility. Paths must be validated, writes must be atomic, and stale sessions must fail clearly. “Local-first” is not permission to be careless with the filesystem.

Human control is a real product boundary

The visual editor is not a preview window for agent output. It is the place where a person can disagree. That means agent instructions must describe the goal and constraints without repeatedly overwriting deliberate edits.

The best automation I have built does not hide its seams. It makes the handoff legible, keeps the source editable, and gives both sides a reliable way back.