Key takeaways
As GPT-6 Astra gets more capable, skills, AGENTS.md, and task prompts piled up for weaker models now bloat context, contradict each other, and mis-trigger. Keep skill descriptions short with precise triggers; use progressive disclosure so the root doc is a router; point AGENTS.md at docs by task instead of full pre-reads; rewrite ask-first walls and define completion up front so Astra’s more tentative style doesn’t stop early. A new model is a good time to clean the instruction house.
