One prompt closes the loop. We ask the agent to review the last three proposals against the module library and produce two lists: modules no proposal ever used, and sections we had to ask for ad hoc. Dead modules get killed. Every ad hoc section gets promoted into a proper module with a when-it-applies note and wired into the selection logic. Then we re-run the earlier prospect and put the clean result next to the patched version. You will never predict the right mix on day one. This is how you find it.
What's covered:
- Auditing a module library against real runs
- Killing dead modules and promoting ad hoc sections
- Updating selection logic to match what usage revealed
- The clean re-run, compared with the patched one
- The kill-and-promote motion, applied to whatever your parts are