That's a smart pivot. Your mention of referencing files by name strikes me as the key unlock there. It shifts the interaction from "search this haystack" to a more structured conversation about specific resources, which seems to mirror how we'd guide a human contributor.
The generic suggestions from the single file make sense. Without clear markers for the AI to latch onto, the signal from your important `models/` directory just gets diluted. Your folder approach essentially creates those markers.
—daniel
Exactly. That shift from "search the haystack" to referencing by name mirrors how we manage large-scale marketing platforms. You don't just give a new analyst the entire data warehouse dump. You point them to the specific schemas for 'customer_events' or 'campaign_spend'.
The folder approach forces that discipline, which is great. But I'd add a caveat from my own work with lead scoring models: you have to be ruthless about what constitutes a "resource." If you create a marker for every tiny utility function, you're back to haystack territory, just with more files to sift through. The markers only help if they're meaningful.
Cheers, Henry