AI in the loop (3.4.0)

As of 3.4.0, when a publisher command like upload or preview fails, you’re one keystroke away from having AI diagnose what went wrong — inside your project, with all the context it needs.

When something breaks

Uploads and builds sometimes fail. Usually it’s a real bug in your code, buried in a wall of build output that’s hard to read and easy to misjudge.

Now, when a publisher command fails, we hand the problem to Claude Code. The publisher writes a diagnostics file — the error, the likely cause pulled from your build logs and pointers to the publisher’s own rules — then offers two ways you can use it with AI:

• Open a Claude Code chat to fix it: The VSCode extension opens a chat right next to your code, pre-filled and ready. It can read your project files and the error diagnostics, find the real cause and fix it with you in the chat. (And it won’t edit files until you confirm you want to.)
• Just tell me what went wrong: Sends the error diagnostics to Claude in your terminal and prints back a quick, read-only synopsis of the likeliest cause — it won’t touch your files.

No copying stack traces into a chat window. No guessing which log line matters. The agent starts with everything the publisher already knows about the failure, scoped to your project — not generic advice.

It’s already on

The prompt appears automatically the next time a command fails in your terminal.

Missed it, or don’t want to re-run a slow build just to look again? You can run the new diagnose command, which will read the error diagnostics from your last failed attempt.

CLI

pnpm graphics-publisher diagnose

Graphics kit

pnpm diagnose

Note

The publisher does not handle setting up or authenticating Claude — it assumes you’ve already done that yourself. Read more about our recommended setup for Claude Code and the VSCode extension in our Notion docs.

This is the first step of bringing AI into the publishing process — a small one, aimed at the moment it’s most useful: when you’re stuck. More to come.