Super Data Science: ML & AI Podcast with Jon Krohn 1022: CLAUDE.md, AGENTS.md, Skills, Hooks and Subagents: A Field Guide to Steering AI Agents
126 snips
Aug 28, 2026 A lively tour of ways to steer AI agents, from where to store instructions to balancing persistence and token cost. Learn about rules, skills, subagents and context isolation. Hear why hooks act as deterministic guardrails and how agents.md is shaping industry standards. Three practical takeaways wrap up how to organize instructions like code.
AI Snips
Chapters
Transcript
Episode notes
Three Properties That Explain All Steering Methods
- Instruction methods differ by when they load, whether they survive long sessions, and how much authority they carry.
- Jon Krohn frames these three properties as the lens to reason about any agent instruction mechanism across platforms.
Context Window Growth Causes Instruction Drift
- A model's context window holds session state and costs tokens and attention as it grows.
- Krohn explains compaction: long sessions get summarized and earlier instructions can be squeezed out, causing rule drift.
Keep Root Instruction Files Short And Owned
- Use a root claw.md (or agents.md) for always-relevant facts like build commands and team norms, but keep it short.
- Krohn advises limiting claw.md to under 200 lines, assign an owner, and review changes like code.