◇ Could this help me?
Yes. Both halves are sound: over-scaffolding a capable model with layers of instructions genuinely does degrade results, and scheduled maintenance passes catch the dead code, missing tests, and duplicated logic that nobody prioritizes by hand. Add one guardrail the talk notes do not mention: a routine that rewrites code while you sleep should produce a reviewable branch or pull request rather than committing to the main line, so every automated change still passes through a human read. This is a general practice.
Set up four recurring maintenance passes for this repository, each producing a reviewable branch rather than committing directly. First, a general maintenance pass for outdated dependencies and broken references. Second, a dead code sweep listing unreachable or unused code with evidence for each finding. Third, a test coverage run that reports the least covered critical paths. Fourth, a duplication pass that finds over-complicated logic repeated across files and proposes a simpler shared version. For each, tell me what it changed and why before I merge anything.