antislop gives coding agents 38 rules against generic UI and copy

The developer miqdadbadjuber has published antislop, an MIT-licensed set of rules that stops AI coding agents from producing generic AI-generated interfaces, copy, and code. The repository was created on 7 August and has picked up 919 stars in the four weeks since.

What it is

antislop is a filter: 38 numbered rules organised in three tiers, plus a final check, the delivery gate, that runs before anything is released. The Hard Gate rules are absolute rejections. The Purpose Gate lets a technique through only when the reason is written down. The Quality Locks enforce consistency across a session. The rules do not prescribe colours, fonts, or layouts; the direction for a piece of work comes from a separate DESIGN.md in the user's own project.

How it fits into an agent

The project comes as standard Agent Skills folders, one per topic, that you copy into the skill directory any compatible tool reads. antislop lists installation paths for Claude Code, Codex, Antigravity, OpenCode, Cursor, Gemini CLI, and Hermes. A single-file version (antislop.md) also works pasted into any chat window that reads Markdown, for agents without a skill folder convention.

The skills

The core skill is the always-loaded filter. Five additional skills load only when a task needs them: antislop-ui for layout, colour, components, motion, and decoration; antislop-copywriting for headlines, calls to action, tone, and the anti-AI-writing patterns that show up in generated text; antislop-human for contrast, keyboard, focus, and states; antislop-layoutmobile for mobile breakpoints, grids, overflow, and tap targets; and antislop-code, added in version 3.1, which removes generic AI-slop comments while leaving the code alone.

Two modes

antislop is used one of two ways. In During mode it guides an agent while it builds, ending with a mandatory pass or fail report in four blocks. In After mode it audits finished work, returns a numbered findings list, waits for the operator to choose which items to fix, then produces a follow-up report. The mode is picked at the start of a session.

The generic look and voice that coding agents produce by default is well known. Bento grids with three cards, dot backgrounds, colored left stripes, fake terminal screenshots that do not show a product, headlines that use the same pattern of three, comments that restate what the code obviously does. Each one is fixable in a code review, but a team merging several agent-generated pull requests a day cannot remove every generic pattern by hand. antislop puts the rejection into the agent's own instructions, so reviewers see fewer of these patterns to begin with. The tradeoff is that the rules are one project's opinions: a team already running an internal design system will have to decide which of antislop's rules to keep and which to override. Real interest shows up in the star count, 919 in the first four weeks, and in the fact that the project has already published installation paths for seven different agents.

Source

Primary source: miqdadbadjuber/anti-slop on GitHub.