What I built: an infographic mode inside my own video skill (reeler). I trigger it with "make an infographic from report X" or /reeler. It reads a markdown report, writes the numbers and short texts into a small JSON, and renders a PNG from HTML in local Chrome. Same idea as Nate's infographic builder, with one change from the start. I need the numbers exact, so the image comes from code instead of an image generator. The AI-generated version is still there on request, but every digit and letter gets checked before it's used. One optimization after watching it work: a guard. Before rendering, a script checks that every number in the JSON exists in the source report. On the first real run it stopped the two totals, because they don't appear in the report as such. Now a total only passes if the sum of its column actually adds up. That check runs as a script, so it doesn't depend on the model getting it right. Six runs so far, and watching them showed me what to fix: an orphan card in the calendar, a sliver on zero-value bars, long names wrapping, and the bottom line cut off because the size was measured before the font loaded. For Dive Deeper I added the MCP part: the same JSON goes to ChatGPT through Claude in Chrome, and the generated image comes back to be checked against it. First try, GPT also chose to build it from code. When I asked for the image generator, every number and word came out right, but the bars were only roughly proportional and the size was 1024×1536 instead of what I asked for. Both images below, same data: code on the left, GPT on the right.