Version 2.3.0 put AI inside the page editor. The person editing describes what they want, and the system returns HTML and CSS ready to insert.
Two kinds of instruction
- Mode: the technical instruction. It tells the AI how the system works, which classes to use and in what format to answer. It is written once and reused.
- Request: what the person wants at that moment, in plain language.
Keeping the two apart means each request does not have to repeat the rules, and the rules can be adjusted without touching saved requests.
What landed
- A library that prepares the instruction, calls the configured AI service and separates the HTML and the CSS in the answer.
- Dedicated tables for connections, modes and saved requests.
- Choice of connection and model when asking.
- Insertion of the result at the chosen position, with the preview updated.
- Error handling when the external service fails.
Generation depends on an AI connection configured by the installation. What the AI returns is a draft: the review still belongs to whoever publishes.
Source: Core · v2.3.0 · commit 94c1ca91 · ia library