Keep exploring with the agent while a new case could change the rules. Hand execution to a script when you can name the inputs, describe what should happen, and check the result.
- Keep exploring when examples still lead to different decisions. Have the agent work through a few real cases and show you where the rules break.
- Write the script when the same rules work across those cases. Give it a check you can run again and a clear error when it hits an input it can’t handle.
Example
For a new Field Guide playbook, work with the agent on the title, description, and placement, then add the entry to the Playbooks index. Run npm run catalogue:sync to rebuild the homepage catalogue from the category pages. Open the homepage on localhost to check how the entry wraps and follow its link.