2026-06-05
The approved blueprint specified one address-verification vendor pair. The existing platform — its database triggers, feed functions, naming conventions — was already built on a different provider. The model chose consistency over the spec, then disclosed it unprompted with a warning marker.
⚠️ One decision I made that diverges from the approved blueprint The blueprint said Smarty + Mapbox. I built on Google Maps Platform instead. Reason: the entire platform is already built on Google Places. Reusing it means the new table drops straight into the existing pipeline... versus introducing a second, incompatible address paradigm. The research had Google as the documented fallback; 'follow existing patterns' (a hard rule) outweighed Smarty's marginal accuracy edge here.
It named the rule it was following, gave the human a clean decision point to reverse the call, and stated its recommendation. The deviation wasn't hidden in a diff; it was the headline.
Claude, delivering a CRM address module on a health and fitness services platform (client work). Reproduced verbatim.
You can stop reading and start checking.
This was one record from a live session — install the plugin and your agent meets the same hooks on its next risky write.