Workflow for processing Jörn's feedback during interactive sessions. Covers the generalization loop — fixing the instance, abstracting the error class, scanning for all instances, and recording durably. Load when receiving corrections or review feedback from Jörn.
During interactive sessions when Jörn flags an error or gives a correction. Not needed for autonomous subagent work.
When Jörn flags a specific error (e.g. "line 15 doesn't check theorem X's preconditions"):
Fix the specific instance. This is the minimum.
Abstract the reasoning error. Not "theorem X on line 15" but "applying results without verifying preconditions." The error class is about the type of reasoning mistake, not the specific content. It won't be the same theorem elsewhere — it will be a different result applied without checking its own preconditions.
Scan for the same error class. Search the current file and other files from this session. Look for the same type of reasoning error, not the same surface pattern. Dispatch subagents for large scopes.
Record the reasoning error in MEMORY.md. Describe the error class with enough context that a future agent in a different session recognizes when it applies. Include: what the error looks like, why it's wrong, and how to check for it.
Re-review. Dispatch a subagent to re-read the deliverable through the lens of the feedback. "Given that I made error X in two places, what else did I miss?"
Jörn's time is the scarcest resource. Every round of feedback he gives should prevent the entire error class, not just the one instance he pointed at. If Jörn has to flag the same type of mistake twice, the feedback loop is broken.