Unfry

Richard Dawkins coined “meme” in 1976 for a unit of culture that replicates, mutates and spreads. He was describing evolution by other means. The definition turned out to fit image macros better than he could have planned for: a format is a container that carries meaning across contexts, and every remix is a unit of culture adapting to a new environment.

A format is recognisable by its structure, its expression and its context, not by the specific words sitting on it. Swap the caption and it is still the same meme. That is the whole point.

Here is the cruel part. The more a format succeeds, the worse it looks. It becomes recognisable because it has been through the cycle, and the cycle is what destroys it. Compressed, rescreenshotted, watermarked by three different sites, cropped by whoever posted it last. By the time a template is famous enough to be worth reusing, it is a ruin. And the text is part of the ruin, sitting in the pixels rather than on a layer, so editing it means covering it over and hoping.

This is not a niche complaint. The same behaviour runs from a stranger's keep-going folder to embassy accounts mid-crisis, and the missing tool is the same at every stop.

Splitting it honestly: I did the product concept, the scope cuts, the name, the design direction and the quality judgement on every output. I also did the one piece of engineering that mattered, which was diagnosing that the doubled text was a sequencing problem and specifying the detect, remove, reapply order. Agent 4 did the frontend, the backend routes, the Gemini integration, the canvas text system, the comparison slider and the responsive layout.

None of it is magic, and it is not automation either: it is a much shorter cycle between judgement and result. Describe the thing, watch it get built, see where it breaks, adjust, go again. The loop that used to take a week takes an afternoon, which moves the constraint off implementation and onto knowing what good looks like.

There is a joke buried in the timing. The discourse that made the problem visible was an argument about action beating overthinking, and the response was to build the thing rather than write a document about building the thing.

What is still rough: the system cannot tell structural text from content text, so it offers you the axis labels on a bell curve as though you might want to rewrite them. Detection drops characters and does not flag when it is unsure. Text rotation and several editor features were specified and never finished. The tool works and it shipped, and that is the whole claim.