Hoard Your Coding Knowledge: The Secret Weapon of AI-Assisted Developers

Hoard Your Coding Knowledge: The Secret Weapon of AI-Assisted Developers

In the rapidly evolving landscape of AI-assisted programming, the most valuable asset a developer can cultivate isn’t a new framework or a cutting-edge tool—it’s a personal repository of proven solutions. According to Simon Willison, a leading voice in agentic engineering, the practice of hoarding things you know how to do has become a foundational strategy for maximizing the utility of coding agents.

Willison, whose extensive public portfolio includes over a thousand GitHub repositories and a curated collection of browser-based tools, argues that deep, hands-on experience with niche technical challenges forms the bedrock of effective AI collaboration. Rather than relying solely on theoretical knowledge or generic documentation, he emphasizes the power of documented, executable code snippets—each a tiny monument to a problem once solved.

The Science Behind the Hoard

Modern coding agents, whether powered by Claude, GPT, or other large language models, excel at synthesis. But they require high-quality input. A model may know that OCR can be performed in JavaScript or that PDFs can be rendered into images via PDF.js—but without concrete, tested examples, it struggles to assemble them into a working application. Willison’s breakthrough insight is that seeing something done is fundamentally different from knowing it’s possible.

His personal hoard includes hundreds of small, self-contained HTML tools hosted at tools.simonwillison.net. These are not prototypes; they are production-grade, browser-native applications—each solving a single, well-defined problem. One such tool, an OCR processor for scanned PDFs, was built by combining two previously documented snippets: one for rendering PDF pages as images using PDF.js, and another for running Tesseract.js to extract text from those images. The result? A fully functional, drag-and-drop web application that took him minutes to generate with an AI agent, but years to accumulate the components for.

Agents as Librarians, Not Inventors

Willison’s workflow illustrates a paradigm shift: coding agents are not replacements for developers—they’re hyper-specialized librarians. By instructing an agent to “fetch the source of https://tools.simonwillison.net/ocr” or “clone simonw/research from GitHub,” he transforms abstract prompts into concrete, executable tasks. The agent doesn’t guess—it retrieves, adapts, and integrates.

He often uses this technique to accelerate testing and documentation. For example, he might tell an agent: “Add mocked HTTP tests to ~/dev/ecosystem/datasette-oauth inspired by how ~/dev/ecosystem/llm-mistral is doing it.” The agent then searches his own codebase, identifies patterns, and applies them—a process far more accurate than generic instruction-following.

Building a Sustainable Knowledge Infrastructure

Willison’s approach is systematic. He uses his personal blog and TIL (Today I Learned) site to capture fleeting insights. Each entry includes not just an explanation, but a runnable example. His GitHub repositories are meticulously organized, with even the smallest proof-of-concept tagged and documented. This infrastructure turns individual learning into institutional knowledge—knowledge that scales with his team, his tools, and his future self.

As AI agents become more integrated into daily development workflows, the ability to curate and retrieve this kind of granular, practical knowledge will become a defining skill. Developers who hoard solutions don’t just work faster—they solve problems others can’t even conceive of, because they’ve already seen them done—and documented.

As Willison puts it: “We only ever need to figure out a useful trick once.” The rest is retrieval, recombination, and refinement—powered by the quiet discipline of keeping a personal library of what works.