The in-editor AI panel in Alex CMS started as a toolbox: twenty-two actions, each a button, each equally weighted. Improve this field. Draft a summary. Check links. Generate SEO. Score the page. All real, all useful, all generated quickly.
Using it made the problem plain. Twenty-two equally weighted buttons is the same as no guidance. The panel answered "what can the AI do?" when the editor's question was "what should I do next?"
The rebuild. A page-health card at the top. One recommended next step, chosen from the page's actual state. Four collapsible stages underneath for everything else. The action labels were rewritten outcome-first ("Make the summary clearer" rather than "improve.summary") as a display-only change, with routing still keyed to stable action keys, so renaming a button can never break the automation behind it.
The lesson. AI made the twenty-two features nearly free to build, which is exactly why the panel filled up. The scarce thing was the judgment about which one to lead with, and that never came from the model. It came from watching someone use it and admitting the first design was wrong.