The evolution of my AGENTS.md- v0 - Braindump- v1 - Organized braindump- v2 - Strict state machine- v3 - Loose state machine, md as memory- v4 - Let s add skills!- v5 - Shorter prose, more skills- v6 - Sectioned. Each with - Bias via short prose - Yaml block of sect structure - Deep dive pointer to skills/[section]/SKILL.md - v7 - ??? - define the std sections Discoveries:- you can NOT instruct llms. You can only introduce bias to token generation.- repo based memory via .md files is the way- frequent session restarts saves both money and agent performance