Simon Willison, a prominent software engineer and co-creator of the Django framework, advocates for a unique productivity strategy: hoarding personal code snippets to enhance the performance of AI coding agents. Instead of relying on an AI's general knowledge, developers should curate a library of small, functional proof-of-concepts. This personal archive acts as a ground truth—a reliable reference that ensures AI systems produce working results based on the developer’s specific, proven methods rather than generic patterns.

The real power of this technique is found in recombination, where a user prompts an AI to fuse multiple existing examples into a new tool. For instance, by providing code for a PDF renderer alongside an image-to-text library, a developer can quickly generate a custom OCR application. This workflow effectively bypasses the AI's tendency to hallucinate incorrect code, as it is constrained by the logic of the user's provided samples.

As modern AI tools gain the ability to index local files and browse the web, the value of maintaining a well-documented knowledge hoard becomes even more apparent. These coding agents—autonomous or semi-autonomous AI systems—can now search a developer's history to find the perfect testing template or architectural pattern for a new project. This transformation means that once a technical hurdle is cleared, it becomes a permanent building block for future AI-assisted creation.

Simon Willison, a prominent software engineer and co-creator of the Django framework, advocates for a unique productivity strategy: hoarding personal code snippets to enhance the performance of AI coding agents. Instead of relying on an AI's general knowledge, developers should curate a library of small, functional proof-of-concepts. This personal archive acts as a ground truth—a reliable reference that ensures AI systems produce working results based on the developer’s specific, proven methods rather than generic patterns.

The real power of this technique is found in recombination, where a user prompts an AI to fuse multiple existing examples into a new tool. For instance, by providing code for a PDF renderer alongside an image-to-text library, a developer can quickly generate a custom OCR application. This workflow effectively bypasses the AI's tendency to hallucinate incorrect code, as it is constrained by the logic of the user's provided samples.

As modern AI tools gain the ability to index local files and browse the web, the value of maintaining a well-documented knowledge hoard becomes even more apparent. These coding agents—autonomous or semi-autonomous AI systems—can now search a developer's history to find the perfect testing template or architectural pattern for a new project. This transformation means that once a technical hurdle is cleared, it becomes a permanent building block for future AI-assisted creation.