Why LLM instructions drown in long context windows

How much weight do you actually expect a model to give to a "never" command once a conversation hits a certain length? I hit this wall while building an agent—I had a clear directive to avoid direct database writes, but as the context window filled up, the model just... ignored it. The instruction was sitting right there in the history, but the attention had decayed. It wasn't a logic error; it was a structural one.

I started thinking about this in terms of topography. If you treat every rule as a "hill" in a 3D space, a well-anchored, imperative command with backticks stands high. A polite or vague request is just a low mound. As the "context load" (the water level) rises, those weak rules sink first. Even high-stakes constraints fail if they are written as prose instead of hard, structural anchors.

To test this, I built a tool that visualizes this decay using a Gaussian height field on a raw 2D canvas. No heavy Three.js overhead—just a painter's algorithm to show you which parts of your prompt are about to drown. I also added a linter that flags hedging or excessive politeness, helping you catch structural weaknesses before they cause a production incident.

If you're building agents, shouldn't a mission-critical rule be a runtime hook rather than just a sentence in a prompt?

You can play with the interactive demo and see the logic behind how instruction quality affects height here:

https://reporails.com/demo/see-why-ai-instructions-decay-then-write-ones-that-hold

The breakdown of how I categorize these rules is here:

https://reporails.com/rules/core

If you want to see how the rendering works or just fork the single HTML file to use in your own testing, check the source:

https://codepen.io/editor/G-bor-M-sz-ros-the-reactor/pen/019f4cad-e344-78bf-b7bc-919972f42a4e