Timon Krebs, an author and software developer, argues that in AI-driven coding tasks, the entire codebase serves as the prompt rather than being separate material for the model to work on. He observes that an AI model does not act as a database of facts, but rather as a system that reconstructs plausible continuations based on the patterns it perceives. When working on a codebase, these patterns include naming conventions, architecture, hidden assumptions, test status, and file layout, all of which function as part of the query.

Krebs details his experience using Fable 5, an AI coding tool, which he found highly effective for reviving stagnant personal projects. He notes that the model was able to solve complex problems by helping him visualize the shape of the solutions. However, he encountered inconsistent performance when applying the tool to professional codebases. This performance variance highlighted that output quality is intrinsically linked to the underlying code structure. Good code, which emphasizes encapsulation and clear interfaces, is local and compresses well, allowing models to infer missing information easily. Conversely, bad code is non-local, relying on hidden global states or distant dependencies that often fall outside the model's context window.

The author explains that simply providing a model with clean examples or "best practices" often fails because these instructions become part of a contradictory query. In a messy codebase, the model may follow the best practice in isolation but fail to account for the implicit system constraints that caused the code to be irregular initially. He describes this as a form of context rot, where stale files and irrelevant examples remain in the model's active window, confusing the output. Ultimately, Krebs concludes that improving performance on messy codebases does not rely on cleverer prompting, but on supplying specific, non-local context. Success with LLMs involves shrinking the task to ensure locality and identifying hidden invariants so the model can function within a defined boundary.

Timon Krebs, an author and software developer, argues that in AI-driven coding tasks, the entire codebase serves as the prompt rather than being separate material for the model to work on. He observes that an AI model does not act as a database of facts, but rather as a system that reconstructs plausible continuations based on the patterns it perceives. When working on a codebase, these patterns include naming conventions, architecture, hidden assumptions, test status, and file layout, all of which function as part of the query.

Krebs details his experience using Fable 5, an AI coding tool, which he found highly effective for reviving stagnant personal projects. He notes that the model was able to solve complex problems by helping him visualize the shape of the solutions. However, he encountered inconsistent performance when applying the tool to professional codebases. This performance variance highlighted that output quality is intrinsically linked to the underlying code structure. Good code, which emphasizes encapsulation and clear interfaces, is local and compresses well, allowing models to infer missing information easily. Conversely, bad code is non-local, relying on hidden global states or distant dependencies that often fall outside the model's context window.

The author explains that simply providing a model with clean examples or "best practices" often fails because these instructions become part of a contradictory query. In a messy codebase, the model may follow the best practice in isolation but fail to account for the implicit system constraints that caused the code to be irregular initially. He describes this as a form of context rot, where stale files and irrelevant examples remain in the model's active window, confusing the output. Ultimately, Krebs concludes that improving performance on messy codebases does not rely on cleverer prompting, but on supplying specific, non-local context. Success with LLMs involves shrinking the task to ensure locality and identifying hidden invariants so the model can function within a defined boundary.