Skip to content

antialgorithm agent skill steers AI past its first idea

A six-day-old MIT-licensed skill makes models state their most average answer, then deliberately move away from it.

Draftship Editorial4 min read

A skill built to dodge the obvious answer

A repository called antialgorithm, an agent skill for creative work, has 553 stars and 33 forks on GitHub as of 2026-11-20. Those numbers arrived in the six days since the project was created on 2026-11-14. The antialgorithm agent skill is written in Python, carries an MIT license, and is tagged for agent skills, Claude Code, Claude skills, Codex, creative writing, creativity and design.

The antialgorithm agent skill targets a familiar failure mode: the model's first suggestion is often the safest and most generic one. Names, copy, story ideas, game ideas, posters and UI are all listed as its territory.

How the skill works

The mechanism the README describes is two steps. The model first has to state its first, most average pick out loud. It then moves away from that pick for a reason that comes from the project itself, rather than from a general instruction to be more creative.

Four Python scripts support the process. seed.py rolls a starting point when the request is thin. A visual seed supplies a making method, colours, layout, type, a verb and scale for the idea, a setting, a region and, half the time, an object. A text seed supplies a voice, a way in and a register. An idea seed supplies a verb, a constraint, a scale and a pressure. Values come from system randomness, so two people sending the same prompt start in different places. Each seed is written to ~/.local/state/antialgorithm/history.jsonl, or under $XDG_STATE_HOME when that is set, and values from the last few seeds are skipped so sessions do not repeat. --keep "axis=value" pins one axis, which lets a subject stay fixed while the look is rerolled.

moves.py keeps a short log of the main move behind each off-default answer, with examples like "neon tube wall, almost no type". The model reads that log before presenting new work, so it does not reuse a move just because it landed last time. rough.py perturbs an SVG, shaking straight lines, shapes and letters and breaking the even spacing of repeated marks, and the README limits it to hand-made methods such as print, paint, stamps and ink. check.py renders an SVG at full size and at icon size and flags text that runs off an edge, sits under a shape, has low contrast or is too small. Its size limit is tuned for posters, and the README says app screens should use --min-text 0.6.

Visual rules sit in visuals.md, which the model reads only for visual work, so naming a product or writing a slogan does not load them.

Installing it

The skills CLI places the skill where an agent looks for them, and the README names Codex, Cursor, OpenCode and Claude Code among the targets: npx skills add mankidev/antialgorithm. Cloning into the agent's skills folder also works. Claude Code can install it as a plugin, and on versions before 2.1.275 the marketplace has to be added first. Python 3 is required for the scripts, and check.py also needs numpy, Pillow and either rsvg-convert or the cairosvg package.

What the demo shows, and what it leaves open

The README compares two sets of SVG posters for a music festival, 400x600, generated from one prompt. The row without the skill is a batch of six runs with none left out. The row with the skill was picked from about a dozen runs made while the skill was still being written. That is an illustration rather than a controlled test, which suggests readers should treat the images as a demonstration of direction rather than a measured improvement. The README gives no scoring method and no list of which models produced the comparison, and the author describes the second row as a pick.

It is also unclear how the skill behaves with agents that do not read skill files, or how much of the difference comes from seed randomness versus the instruction to move away from the first pick. The repository is six days old.

For people building apps

Most of the examples are visual, which matters for anyone generating interface layouts or marketing assets with a model. The check.py flag for text under a shape or below a contrast threshold is the closest thing here to an app-screen check, and the README is explicit that its default size limit is set for posters, with --min-text 0.6 suggested for app screens. The seed history file is also worth knowing about: the README says plugin updates leave it alone, so repeated prompts drift instead of repeating.

For a builder working in Draftship, the transferable part is the framing: asking a model to write down its predictable answer before producing anything, and then pushing against that answer with a constraint drawn from the project. Writing effective prompts covers how to give a model that kind of context. Draftship is not affiliated with this project and does not ship the skill.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.