Error-Driven Prompt Tuning Loop

Build a loop that collects model errors/failures and proposes prompt edits. Steps: (1) bucket failures, (2) extract missing instructions, (3) propose deltas, (4) generate 3 patched prompts, (5) quick re-test table. Output as YAML with sections: failure_buckets, root_causes, proposed_deltas, candidates[], retest_plan.