An AI's First Idea Is Always the Most Boring. This Skill Forces It to Step Away
What This Thing Actually Does
Ever notice how you ask an AI to name your product, write a tagline, or design a poster, and the very first answer it gives you is always the safest, most forgettable option? Not wrong, just... average. antialgorithm is an Agent Skill built specifically to fix that. The core mechanism, in plain terms, has two steps: first it makes the model state out loud the first and most average pick sitting in its head, then it requires the model to move away from that spot for a reason drawn from the project itself. There is a bell-curve chart in the README that makes this click instantly. Most choices pile up around the middle peak, which the project calls the safe zone, and this skill pushes you out toward the tail. Names, stories, game ideas, logos, posters, and UI all count as creative work here. Install it once and your agent gains a switch that says, essentially, don't hand me the most ordinary answer.
Why It Blew Up Recently
This repo was only created in mid-November, yet it picked up a wave of attention within days, and I think two things drove it. One is timing: Claude Code, Codex, and Cursor all started supporting skills folders this year, and Vercel even shipped a dedicated skills CLI, so installing a skill became as easy as installing an npm package. The ecosystem caught fire almost overnight. The other is that it hit a pain point everyone feels but nobody had tackled systematically, namely that AI creative output reads like correct-but-bland filler. What really carried it across social platforms was the showcase comparison image: six music-festival posters from the same prompt, the top row generated without the skill and all crowded together looking nearly identical, the bottom row generated with it and each visibly distinct. The difference is obvious at a glance. Ten translated READMEs lowered the barrier even further, so the recent star growth makes total sense.
The Interesting Bits
The genuinely interesting part is the four Python scripts tucked into the scripts folder. seed.py rolls a starting point when your request is thin. A visual seed hands you a making method, colors, layout, type, a verb, a scale for the idea, a setting, a region, and half the time a concrete object. A text seed gives you a voice, a way in, and a register. An idea seed gives you a verb, a constraint, a scale, and a pressure. It draws from system randomness, so two people sending the exact same prompt start from completely different places. Even better, it saves each seed to ~/.local/state/antialgorithm/history.jsonl and skips recent values, so you do not repeat yourself across sessions, and plugin updates leave that file untouched. Want to hold one axis fixed? Use --keep "axis=value", for example keeping the subject and rolling only the look.
moves.py is a move log. It records the main move behind each off-default answer, something like neon tube wall with almost no type, and the model reads it before showing new work so it will not reuse a move just because it worked last time. rough.py fixes the problem of code-drawn art looking too plotted: it shakes the straight lines, shapes, and letters of an SVG and breaks the even spacing of repeated marks, so hand-made methods like print, paint, stamps, and ink actually look hand-made. Screen, camera, and render looks are left alone. check.py is the quality inspector. It renders an SVG at full size and at icon size, then lists text running off an edge, text sitting under a shape, low contrast, and type that is too small. Posters use the default limits; for an app screen you reach for --min-text 0.6. There is also a visuals.md that the model reads only for visual work, so a name or a slogan does not load it. The attention to detail is impressive.
Who It Is For
If you are a designer, a brand or naming professional, an indie developer, or anyone doing creative work with AI who is tired of the same-old output, this skill is basically made for you. It does not pretend to dream up something brilliant on your behalf. It simply puts the most mediocre option on the table and then kicks it away, giving your inspiration a fork in the road. People studying prompt engineering should also look at the approach, because it turns the abstract idea of going against the default into a concrete, reproducible set of scripts and constraints.
Quick Start
The one-liner is npx skills add mankidev/antialgorithm, which drops the skill wherever your agents read skills. You can also git clone it into your skills folder, and Claude Code users can install it as a plugin. The scripts need Python 3, and check.py additionally needs numpy, Pillow, and rsvg-convert or the cairosvg package.
How It Compares
- Adding "be creative" to a normal prompt: barely helps, the model still drifts back to the mean and you cannot steer where it goes.
- Generic creative skills: mostly slogans, lacking reproducible random seeds and deduplication, so a few runs later you are still colliding with yourself.
- This skill: seed.py random axes plus the moves.py log plus history deduplication turn stepping away from the average into a repeatable process with memory, split into visual, text, and idea starting points.
- The catch: it only handles the opening move of creative divergence, not final quality, so poster layout still leans on check.py and your own eyes.