antislop: a filter for AI coding agents, not a style guide

The problem antislop names: generic output that is clean and dead

AI coding agents converge. Ask one for a landing page and you get a sparkle logo, a beta pill, a fake terminal reporting an invented latency figure. Ask for a launch post and you get emoji bullets and hype. The README shows this directly in its comparison assets: the build with nothing applied has a NEXT-GEN AI 2.0 beta pill and a terminal reading 0.0001ms latency, and the same page under DESIGN.md alone still carries stat cards reading 10,000% ROI Synergy Multiplier and a 5.0 rating from 500,000 founders. Direction and filtering are separate jobs, and the second one is what antislop takes on. The audience is anyone whose agent produces the UI, the copy and the code: a frontend engineer reviewing agent-generated markup, a technical writer whose drafts all sound the same, a solo developer who does not want to hand-check every constant for a comment that restates it. The README frames the tool as "a filter, not a style guide": no prescribed colors, fonts or layouts. That restraint is the design decision worth noticing, because it is also the source of its main limitation.

38 rules in three tiers, and a Liveliness Toolkit with three dials

The rules live in the rules/ directory and run from R-01 to R-38, split into three tiers. The Hard Gate is absolute. The Purpose-Gate allows a technique but requires a stated reason for it, which is how a pattern like a gradient or an animated counter survives review: the agent has to justify it rather than default to it. Quality Locks handle consistency across a build. On top of the rules sits the Liveliness Toolkit, with three dials named ENERGY, RHYTHM and MOTION, plus a Design Read. The point of those dials is the failure mode a pure filter would otherwise create. A rule set that only removes things converges on the same beige result every time, so the toolkit pushes back toward something alive and specific. R-37 states the boundary plainly: the core prevents slop but cannot invent direction, and a sterile result means the direction was missing rather than the filter having failed. Skills are additive, one per concern, so an agent loads only what a task needs. The core is always loaded.

Installing antislop and running a first build

antislop ships as standard agent skills: one folder per skill, each holding a SKILL.md. The README points new users at GUIDE.md first, and describes five install routes. The recommended one is the installer, which asks which extra skills you want, whether to install for this project or everywhere, and which agents you use, then copies the folders.

npx antislop-ai

On a project install, the installer is also the only route that writes the pointer which reloads antislop every session. A global install relies on the skills loading themselves by description. To update later, run the same command and choose Overwrite them when it finds the existing folders; choosing Keep what is there installs nothing and leaves you on the old version.

The second route goes through skills.sh, the open directory for agent skills, and copies the same folders without the pointer:

npx skills add miqdadbadjuber/anti-slop

Claude Code users add the marketplace once and install the plugin:

/plugin marketplace add https://github.com/miqdadbadjuber/anti-slop
/plugin install antislop@anti-slop

Antigravity has its own CLI route, and Codex has a marketplace plus plugin route. For a first real use, the README's own comparison is the clearest test: give your agent a small landing page brief, then check the output for the specific tells it lists, such as invented numbers, a beta pill or a fake terminal. Before the work ships, the Delivery Gate expects a mandatory PASS/FAIL report in four blocks. If that report does not appear, the skill is not loading, and the install route you chose is the first thing to check.

What antislop will not do for you

The README is unusually direct about this: the core cannot invent direction, and a sterile result means the direction was missing. If you install antislop, apply no DESIGN.md, and ask for a marketing page, you should expect honest but plain output. The README's own comparison shows exactly that state in the antislop-alone column: honest copy on a restrained dark layout with a single accent color, described as plain because beauty is not the filter's job. That is a real cost, not a footnote. The tool also does not audit anything on its own initiative. The README says it "never beautifies on its own" and that DESIGN.md is where beauty and direction come from, so the workflow only works if a human writes that file. Teams that want a single command to produce a finished, on-brand page will find antislop removes more than it adds. The comparison table makes the pairing explicit: only the build with both antislop and DESIGN.md is clean and directed at the same time.

How antislop differs from a linter or a design system

A CSS linter checks syntax and property usage. A design system hands you tokens, components and spacing scales. antislop does neither. It is a set of instructions an agent reads before it generates, and the rules are phrased as gates the agent must pass rather than properties a tool can verify. That is why the Delivery Gate is a written PASS/FAIL report instead of a failing exit code: the check happens inside the agent's reasoning, not in a build step. The practical difference shows up in what each approach can catch. A linter will never flag a stat card reading 10,000% ROI Synergy Multiplier, because the markup is valid. A design system will not flag it either, because the card uses the right tokens. antislop's Hard Gate is aimed at exactly that class of output. The trade-off is verification: because the rules run as agent instructions, you cannot run them in CI and get a deterministic result. The README's own framing of the tool as a filter rather than a style guide is the honest description of where it sits between those two alternatives.

Maintenance, licence and upgrade cost

The repository is MIT licensed, which permits commercial use and modification, with the usual absence of warranty; that is a description of the licence text, not legal advice. The last push was on 2026-09-16, and the release history shows rapid iteration: v3.2.7 on 2026-09-11 added an OpenCode route and named three more slop patterns, v3.2.8 on 2026-09-14 covered Antigravity's global path, the entry-file writer and a silent exit 0, and v3.2.9 on 2026-09-15 gave Hermes a project scope and rewrote the install documentation end to end. That cadence is the upgrade cost. Rule identifiers are stable enough to cite (R-01 to R-38), but install routes and agent coverage have changed within days of each other, so the version you installed last month may not match the current README. Updating means running the installer again and choosing Overwrite them; choosing Keep what is there installs nothing and leaves you on the old version. Since the skills are plain folders holding SKILL.md files, you can read the diff before overwriting.

Editorial conclusion

Adopt antislop if an agent writes your UI, copy or code and the output keeps landing on the same invented numbers, sparkle logos and emoji bullets, and if you can write a DESIGN.md to fill the space the filter clears. Skip it if you expect the tool to make a page beautiful on its own: the README states that a sterile result means the direction was missing, not that the filter failed (R-37). Before you commit, run npx antislop-ai in one project, check which skills it copies and whether it writes the pointer file that reloads antislop every session, then confirm the four-block Delivery Gate report actually appears in your agent's output.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. miqdadbadjuber/anti-slop on GitHub
4. README
5. Releases