Behind the scenes stuff. This week most of the work went into ImageAI's Layout tab, and we rebuilt it from the ground up.

The old Layout tab was a template picker. Pick a template, hope it fits, move on. We replaced that whole idea with something we're calling the AI Layout Designer, and it follows a simple loop: describe a page, let the AI design it, iterate, then fill it. It shipped in pieces (eight PRs, #18 through #25) because the thing is big and we wanted each phase to have its own spec, plan, and review before it landed.

Here's roughly how it came together:

• Phase 1 laid the foundation, the structural plumbing everything else stands on.
• Phase 2 made the tab design itself. You describe a page or type a change, the text model proposes a layout (or asks you a clarifying question), and it renders on the canvas. Every iteration gets saved to a browsable history you can snapshot, restore, or branch.
• Phase 3 added a per-project style system (named font roles and a palette) plus shareable templates.
• Phase 4 was the Content MVP. The designer produces placeholders, and now you can fill them with real images and text, then export the finished page to PDF.
• Phase 5a brought AI prompt help and .iaibundle export/import, so you can pack up a layout and hand it to someone else.
• Phase 5b wired the Layout tab to the rest of the app: send a design straight to the Image tab, a layout-complete mode, and the core of batch-fill.

We also fixed four user-reported UX papercuts on the way. The big one: the LLM status console now starts collapsed behind a "Show designer output" toggle. It still auto-expands the moment something fails, because hiding an LLM error is the one thing we won't do.

And we made AGENTS.md the canonical coding guide. Every AI coding CLI we use (Claude Code, Codex, Copilot, Gemini, Pi) now reads the same file. The new boundary is simple: coding is full latitude, admin and infra is hands-off. CLAUDE.md and GEMINI.md just import it.

That's the week. The Layout Designer is the piece we're most excited about. Have you ever wanted to describe a page and watch it build itself? Give it a try and tell us where it breaks.