Gábor Mészáros developed an interactive visualization tool to demonstrate the invisible failure of AI models when instructions decay due to high context load. The project illustrates that as an agent’s context window fills, model attention weakens, causing previously defined rules to lose their effectiveness. In this simulation, rules are represented as hills with heights determined by the quality of their phrasing; poorly written or vague instructions sit lower and submerge faster as the 'water' of context load rises. High-stakes prohibitions, such as 'Never write directly to the database,' are highlighted in red to demonstrate that critical constraints are not immune to this attention-fade and should instead be implemented as code-based runtime hooks.

The demo features nine instruction patterns and an integrated 'instruction coach' that uses client-side linting to analyze syntax. The coach identifies common weaknesses, such as politeness, hedging, or improper ordering, where a ban placed before a directive might accidentally prime the model to execute the forbidden action. Mészáros built the 3D field using a 2D canvas with hand-coded projection and Lambert shading, intentionally avoiding third-party libraries like WebGL or three.js. The hill heights are pre-assigned based on human review to ensure consistency, rather than being recomputed dynamically, which the author argues avoids problematic drift in the visualization.

The project emphasizes the distinction between instruction-following in prose versus execution via code. By treating attention-fade as a measurable property, the tool aims to make agent behavior more predictable and empirical. The entire application is a self-contained HTML file without back-end infrastructure or external dependencies, reinforcing the concept that these prompt-based failures occur within the model context itself. The source code is publicly accessible on Codepen and the live tool is hosted on the Reporails platform, providing developers with a hands-on way to test instruction durability under varying context pressure.

Gábor Mészáros developed an interactive visualization tool to demonstrate the invisible failure of AI models when instructions decay due to high context load. The project illustrates that as an agent’s context window fills, model attention weakens, causing previously defined rules to lose their effectiveness. In this simulation, rules are represented as hills with heights determined by the quality of their phrasing; poorly written or vague instructions sit lower and submerge faster as the 'water' of context load rises. High-stakes prohibitions, such as 'Never write directly to the database,' are highlighted in red to demonstrate that critical constraints are not immune to this attention-fade and should instead be implemented as code-based runtime hooks.

The demo features nine instruction patterns and an integrated 'instruction coach' that uses client-side linting to analyze syntax. The coach identifies common weaknesses, such as politeness, hedging, or improper ordering, where a ban placed before a directive might accidentally prime the model to execute the forbidden action. Mészáros built the 3D field using a 2D canvas with hand-coded projection and Lambert shading, intentionally avoiding third-party libraries like WebGL or three.js. The hill heights are pre-assigned based on human review to ensure consistency, rather than being recomputed dynamically, which the author argues avoids problematic drift in the visualization.

The project emphasizes the distinction between instruction-following in prose versus execution via code. By treating attention-fade as a measurable property, the tool aims to make agent behavior more predictable and empirical. The entire application is a self-contained HTML file without back-end infrastructure or external dependencies, reinforcing the concept that these prompt-based failures occur within the model context itself. The source code is publicly accessible on Codepen and the live tool is hosted on the Reporails platform, providing developers with a hands-on way to test instruction durability under varying context pressure.