Self-Learning Image Tagging Loop Each cycle: Agent tags → Human corrects → Corrections become rules + examples → Agent gets better REPEATING CYCLE corrections + explanations GETS BETTER ↑ 📁 New Images Batch of 10 files dropped into folder on shared drive No tags yet 🤖 Agent Builds prompt from: • rules.md • training-set.json Sends to Vision AI 🏷️ Tagged Files XMP keywords written into each file via exiftool 👤 You Review Open review file Edit / add / remove tags Add note: why you changed each tag 📜 rules.md Grows each cycle — distinctions + fixes 📚 training-set.json Human-verified pairs. Grows every cycle. ⚗️ Distill Agent reads corrections Extracts reusable rules Adds verified examples 📋 corrections-log.json Raw record of every change "tag X → tag Y because..." Permanent audit trail QUALITY Cycle 1 Cycle 2 Cycle 3 Cycle 4 Cycle N How the learning works ① Seed (Cycle 1) You hand-tag 10 images. Added to training-set as seed examples. ② Tag + Review (every cycle) Agent tags 10 new images. You correct + explain why. Explanations drive rules. ③ Distill (automatic) Agent reads corrections, extracts rules, adds verified examples. ④ Compounding improvement Richer rules + more examples = fewer corrections needed each cycle.