# How to Automate Continuous Training for AI Agent Skills

Trey, a senior software engineer at Microsoft, explains how to treat agent skill files such as skill.md, claude.md, or agents.md not as static documentation but as trainable, portable software that holds business logic. His central idea is to improve these markdown skills with an automated loop: score the current skill on batches of tasks, have a separate model propose a small edit based on failures, rerun evaluation on held-out tasks, and accept the change only if paired scoring shows the new skill clearly performs better; otherwise, a git reset discards it. He argues this measured process is safer than manually patching prompts after bad outputs, because skills can rot, hidden regressions are common, and model randomness can make simple averages misleading. Using SQL and skill-review examples, he shows that bounded edits, cheap checks before expensive ones, held-out scoring, separate judges and editors, and remembering rejected edits can improve skills significantly without adding inference-time cost.