antialgorithm

Claude Code skill for creative work: names, copy, story, game ideas, posters, UI. It makes the model name its first, most average pick, then move away from it for a reason taken from the project.
Both rows came from the same prompt: Design a poster for a music festival as a 400x600 SVG. Save the final SVG as final.svg in the current folder. The top row is one batch of six runs without the skill, none left out. The bottom row is picked from about a dozen runs made while the skill was still being written.

Install
git clone https://github.com/mankidev/antialgorithm ~/.claude/skills/antialgorithm
The scripts need Python 3. check.py also needs numpy, Pillow and rsvg-convert (or the cairosvg package).
Scripts
seed.py rolls a starting point when the request is thin. A visual seed gives a making method, colors, layout, type, a verb and a scale for the idea, a setting, a region and, half the time, an object. A text seed gives a voice, a way in and a register. An idea seed gives a verb, a constraint, a scale and a pressure. It draws from system randomness, so two people sending the same prompt start in different places. It also saves each seed to history.jsonl in the skill folder and skips values from the last few, so your own seeds don't repeat from one session to the next. --keep "axis=value" holds an axis fixed, which lets you keep a subject and roll only the look.
moves.py keeps a short log of the main move behind each off-default answer, such as "neon tube wall, almost no type". The model reads it before showing new work, so it doesn't reuse a move that worked last time just because it worked.
rough.py takes an SVG and shakes its straight lines, shapes and letters, and breaks the even spacing of repeated marks, so a hand-made look drawn in code stops looking plotted. It is only for hand-made methods such as print, paint, stamps and ink; screen, camera and render looks keep their own marks.
check.py renders an SVG at full size and at icon size and lists text that runs off an edge, sits under a shape, has low contrast or is too small. The size limit fits posters; for an app screen use --min-text 0.6.
Visual rules live in visuals.md, which the model reads only for visual work, so a name or a slogan doesn't load them.
License
MIT