There’s a simple way to stop fixing the same AI mistakes every day. Agents can read a rules file on each run. But most people just fix the output in the chat window and move on. A better way I’ve found is turning every correction into a line in rules(.)md, so the skill always reads it. There’s a key here that’s easy to miss… Don’t just fix the output you got back, also fix the check that should have caught it. How to build these rules: Option 1: Write them yourself every batch. → “Never use this font” → “Always add logo in bottom left.” Option 2: Go deeper with your own sessions. → Have it read your last 5 days of sessions → Review comments, rejected drafts, edits, etc → Prompt it to list instructions you keep repeating Now save them into rules(.)md. This way, whenever your agent generates the next batch: 1. The skill reads rules(.)md before it starts 2. The judge checks against your rules & rubric 3. The mistakes get caught autonomously With this, you’ll stop correcting and only need to review the batch.