Agent-Driven Recombination (Examples + Prompt => Integrated Prototype)

Issue 57 Edition 2026-02-26 6 min read

General

Sources: 1 • Confidence: Medium • Updated: 2026-04-12 10:24

Key takeaways

Collected working examples can be used as inputs for coding agents by prompting agents to build new things by combining two or more existing examples.
Coding agents can fetch and reuse source code from the internet and local codebases as context for new tasks, making the hoarding approach more powerful.
Willison publishes notes in blog/TIL posts and maintains over a thousand GitHub repositories, many of them small proof-of-concepts, as a way to hoard solutions.
tools.simonwillison.net is Willison’s largest collection of LLM-assisted HTML tools implemented as single HTML files with embedded JavaScript and CSS.
Building software skill depends heavily on knowing what is possible and having a rough idea of how to accomplish it.

Collected working examples can be used as inputs for coding agents by prompting agents to build new things by combining two or more existing examples.
Willison built a browser-based OCR tool by combining prior snippets using PDF.js to render PDF pages as images and Tesseract.js (WebAssembly) to extract text from those images in JavaScript.
A single prompt that included both example codebases produced a working proof-of-concept HTML page that converted dropped PDFs into per-page JPEGs and ran OCR with results displayed beneath each image.
Documenting a useful trick once with a working code example enables agents to consult it and solve similarly shaped projects in the future without re-deriving the technique.
Iterating a few times with a model can quickly turn an LLM-generated proof-of-concept into a useful tool with lasting personal value.

Coding agents can fetch and reuse source code from the internet and local codebases as context for new tasks, making the hoarding approach more powerful.
Willison often instructs agents to clone his public repositories to /tmp and mine them for examples to apply to new projects.
When using Claude Code, it can be necessary to specify curl because the default WebFetch tool summarizes pages instead of returning raw HTML.

Willison publishes notes in blog/TIL posts and maintains over a thousand GitHub repositories, many of them small proof-of-concepts, as a way to hoard solutions.
Knowing something is theoretically possible is not equivalent to having personally seen it working as running code.

tools.simonwillison.net is Willison’s largest collection of LLM-assisted HTML tools implemented as single HTML files with embedded JavaScript and CSS.

What is the measurable impact of this hoarding + agent recombination workflow on delivery time, defect rates, or rework versus a baseline workflow without such a repository of examples?
How often do agent-built recombinations fail due to hidden coupling, version drift, or environment/tooling mismatches, and what mitigation practices are required?
Which classes of tasks benefit most from single-file HTML tools, and where does this artifact format become a bottleneck (performance, maintainability, security)?
To what extent is the stated need to use curl in Claude Code a stable constraint versus a transient tool behavior, and what other retrieval-fidelity constraints exist?
Is there any direct decision-readthrough (operator, product, or investor) explicitly supported by the corpus beyond individual workflow suggestions?

If example hoarding plus agent recombination measurably speeds prototyping, demand could rise for developer tools that index, retrieve, and package internal code examples for agents.
If retrieval fidelity and context acquisition dominate outcomes, infrastructure that provides raw, stable, reproducible context for coding agents could see increased adoption.
If single file HTML utilities remain a preferred portability pattern, lightweight tooling that safely generates, audits, and shares such artifacts could benefit.

Controlled comparisons showing reduced delivery time, rework, or defect rates when teams maintain curated runnable examples and use agents to recombine them.
Documented reduction in agent failures after standardizing context retrieval, such as cloning repos locally or providing raw source instead of summaries.
Clear task categories where single file HTML tools outperform alternatives, with repeat use and iteration indicating durability and usefulness.

Studies show no meaningful productivity or quality improvement versus baseline workflows without large example repositories.
High failure rates from hidden coupling, version drift, or environment mismatches persist despite mitigation, making recombination unreliable.
Single file HTML artifacts repeatedly hit security, maintainability, or performance bottlenecks, driving abandonment of the portability pattern.