Duncan is building a website for a tattoo artist friend. Instead of reaching for a vaguely tattoo-looking typeface, he used something the artist had already made: a painted sign from a few years earlier.
Then he asked Codex to turn it into a font.
His thread documents the process in enough detail to be useful. The sign becomes the visual reference. Codex generates the required character set, but the first sheet produces rough edges. Duncan asks for larger groups of characters on a green background instead, making them easier to isolate cleanly.
The useful part happens after image generation
Codex then writes the Python pipeline that builds the font.
According to Duncan, the script uses fontTools, Pillow, NumPy and scikit-image. It keys out the green background, derives an alpha channel, applies a small Gaussian blur to reduce edge chatter, detects rows and characters through pixel projections, traces their outlines, and writes those outlines as TrueType glyphs. The final font is also converted to WOFF2 for the website.
Duncan says the process took about ten minutes. He also says the result could still use cleanup. That distinction matters. This is not a machine replacing type design. It is an automated route from a piece of existing visual language to something usable in a small web project.
That is what makes the experiment worth stealing.
The AI is not being asked to invent a random brand identity. It is doing the boring translation work between a physical artifact and a digital system. For tattooers, sign painters, craftspeople or tiny studios with hand-drawn lettering already scattered across their work, that is a much better starting point than asking a model to invent “a cool custom font” from nothing.