From Context Engineering to Cognitive Engineering
Most prompt engineering treats AI models like people who just need clearer instructions. That assumption breaks everything. Models don't know what they don't know, they can't tell you when an instruction contradicts how they actually generate output, and they'll confidently do the wrong thing if you ask in a way that sounds right but isn't. Model Theory of Mind is a framework for designing prompts based on what the model can actually do, not what you wish it understood. I used it to rewrite Anthropic's frontend design skill and beat the original 75% of the time in blind testing.
Model Theory of Mind starts with a practical question: what can a model actually know, perceive, remember, and act on from the position it occupies? The phrase borrows from theory of mind, but the work is less mystical than that. It is a discipline for not asking models to do impossible things.
The idea came from prompt instructions that sounded reasonable but were impossible for the model to follow, like asking Claude to avoid design patterns it had used in past sessions. The model cannot inspect its own cross-session history. But you can ask it to explore multiple options inside the current generation before choosing one.
That distinction matters. Good prompting means matching the task to the model's real epistemic state instead of the human-shaped one we imagine. The more agentic a workflow becomes, the more expensive those mistaken assumptions get.