Prompts can't add missing information

In response to a post by @theonejvo.

This is related to information theory.

You can’t add information to a system that doesn’t have it.

If you want the LLM to automatically improve your prompts, they can only focus on what you’ve stated from their general view of the world and what a “good” prompt or agentic process looks like — according to static training data.

You can’t magically add information that’s not there. Meaning they can only truly auto improve against your goals if you have given them the connections to do so — your architecture, your UX research, your verification systems.

This is called “Data Processing Inequality” from Cover & Thomas’s Elements of Information Theory: “no clever manipulation of the data can improve the inferences that can be made from the data.”

in other words, you can’t increase information about a target by any kind of processing, only preserve or lose it.

an auto-optimized prompt is just a function of Your information plus the model’s baked in information. The model can add information, but only from typical requests, it can’t know what you might specifically want.

Many people assume that you can prompt less and less as models get smarter but I don’t really find this to be true because they only know what they think matches what a “good” prompt looks like. They don’t actually have training data on interacting with LLMs - yet.

And even if they did, you have to context engineer in order to selectively bring in the information YOU want (system architecture principles you prefer, testing principles important to you, user research, design guidance, etc).