github-code

Drop a `` tag around a normal github.com file link with a `#L9-L18` fragment and it renders just those lines, with line numbers and no syntax highlighting. Under the hood it rewrites the link to the raw-content URL, fetches it, and shows only the requested range — a genuinely useful thing for anyone who writes technical posts or docs and wants to quote live code instead of pasting a stale snapshot.

The tool is small. What's worth reading is packaged next to it: the complete transcript of GPT-5.5 building it from a single conversational message. Willison's whole spec was one paragraph ending in 'Show me a preview web browser so I can see your work' — and the model did exactly that, spinning up a local preview server and iterating against what it could see.

The transcript's tell is what took the effort. The fetch-and-render logic worked almost immediately; the rest of the session is three rounds of cosmetic polish — tightening the header, fixing a full-height separator between the line-number gutter and the code. The hard part wasn't the code, it was the pixels, and the model could only get them right once it could look at its own output.

This is a minor personal tool with essentially no discourse around it. Its value is as a clean, concrete instance of a workflow Willison keeps demonstrating: ship the throwaway tool and the exact prompt that produced it, and let the model close its own visual feedback loop rather than working blind.