About

Recipes going back generations, rebuilt with an AI in the open.

robrambles.com is a personal recipe collection — 178 recipes, some handed down for more than fifty years, curated online since 2016, cooked repeatedly, and finally written down in a form that holds up: exact quantities, real scaling, and no ten paragraphs of preamble before the ingredients.

What makes it worth visiting

Recipes scale properly. Every one is pre-rendered at seven sizes from a quarter batch to a quadruple, in exact fractions rather than decimals that drift. Switching between them is instant and works with JavaScript turned off, because the arithmetic happened when the page was built, not when you loaded it. Where a rounded figure is genuinely approximate, it says so.

Metric sits beside imperial. Where the ingredient's density is known, a cup of flour also tells you what it weighs — which is the measurement that actually reproduces.

Nothing assumes you already know. 377 ingredients and 28 techniques have their own pages, linked automatically wherever an instruction mentions them. What it is, what to look for when buying it, what to use instead.

Unfinished recipes are visible, not hidden. 11 are marked as still being tested. They render in full, with a banner saying so. "Visibly unfinished" and "hidden" are different things, and pretending a half-tested recipe doesn't exist helps nobody.

The AI story

The site used to run on WordPress. In August 2026 the whole thing — a decade of posts — was migrated to a statically generated site with no database and no server-side code, served straight from the edge. That migration, and most of the design and code around it, was done through an extended collaboration with Claude Code.

Not "an AI wrote a website." The recipe data was already structured, so the migration could preserve fractions, ingredient groups, and timings exactly rather than re-parsing prose back out of HTML. Getting there meant auditing the actual export first — which turned up a correlation mechanism that only worked for a quarter of the recipes, one recipe that existed only as fallback HTML, and two recipe cards whose titles had gone stale years earlier. Every one of those was found by building a script and running it against the real data, rather than reading the schema and reasoning about it.

A clean build proves the code compiles. It does not prove a feature produces the right output.

The lesson that recurred in four separate phases of the rebuild

The design decisions are written down as they were made, including the ones that were later reversed and why. So is the work still queued. The site keeps being built the same way it was migrated: a person deciding what it should do, and an AI doing the parts that are mechanical, with everything requiring judgment surfaced as a question rather than guessed at.

Which means this page will go out of date, and the fix will be the same as it has been all along — an edit, a build, and a git push.