A ten-minute conference talk by doron-bleiberg (AWS solutions architect, and bootstrapped co-founder of a coding agent for Minecraft plugins). Common mistakes: prohibitions (“don’t”, “never”), which put attention on the unwanted thing; corrective if-this-then-that patches added after failures; context rot in long-running agents; and re-teaching the model knowledge it already has. The paradox: we want to give the agent everything, but every token competes for attention, costs money, and slows it down. Four principles: evaluate what the model already knows (run several times; include only what it gets wrong); attention pull (say what to do, e.g. “output Markdown”, “answer from the user’s point of view”; avoid implicit hints in eval questions); prefer prevention over correction; and progressive disclosure (minimal system prompt; load task instructions, skills, knowledge, and tool details only when needed, even via hooks at tool invocation).
Show: startup-for-startup
Concepts: context-engineering · skill-engineering · llm-evals