A page with a dark background, oversized headlines and vague promises can look impressive while saying very little about the product. Anti-Slop starts from that frustration. But “AI output looks alike because nothing stops it” is not an established explanation of model behavior. A tool introduction is not research into why generated outputs resemble one another.
Reading the repository revealed a more specific issue: the creator’s final “after” example combines Anti-Slop with design direction. Removing clichés and choosing a visual identity are separate jobs. The project README makes that distinction explicit.
What does Anti-Slop install?
The project is a set of written instructions for coding agents, released under the MIT license. Its current core contains 40 rules, with additional skills for UI, copywriting, accessibility, mobile layouts and code comments. The agent is meant to read the parts relevant to its task. Skill list
“Mandatory” means the instructions tell the agent to follow the rules and produce a PASS/FAIL report before delivery. That report is the agent’s own assessment. Its existence alone does not prove that every check was performed correctly. Delivery Gate definition
What does it reject, and what does it leave to you?
The rules are not all alike. Some impose direct prohibitions: unsourced statistics under R-17, fabricated claims under R-36. Others require a reason for a choice. R-01 allows gradients when they serve an established brand identity or visual hierarchy. “It will never produce gradients or rounded cards again” is therefore an inaccurate reading. Core rules
In our view, the practical value is turning vague feedback such as “this looks too AI-generated” into questions someone can check: Where did this number come from? What does this button do? Does this effect help people read the page? Answering requires inspecting the output.
Nor is every rule a scientific test for machine-generated content. R-02, for example, restricts the “—” character. That is an editorial choice made by the project; seeing that character in a text does not establish that AI wrote it. R-02
What do the before-and-after images actually show?
The following three images were published by the creator in the repository. They are not an independent experiment conducted by us.
Without Anti-Slop: a broad promise in the headline and a terminal displaying an illustrative latency number. Original image
With Anti-Slop, without DESIGN.md. Original image
With Anti-Slop and DESIGN.md. Changes to illustration and composition cannot be attributed solely to the filter. Original image
The second and third images still say “38 rules”; the core file we examined now runs through R-40. These sample images do not match the current text and have deliberately been left unmodified.
Image source: Anti-Slop repository, Copyright (c) 2026 Miqdad Badjuber (antislop), under the MIT license and its terms.
The images help illustrate visual differences. They do not establish a success rate or measure output quality across other projects. This review examines documentation, installer code and published images. We have not run the tool on a project.
Why removing clichés is not enough
R-37 requires design direction from DESIGN.md or explicit guidance from the product owner. Work without direction must be labeled a draft. The file is meant to establish identity, typography, color and mood; Anti-Slop does not make those decisions for you. Design direction requirement
For example, “professional and modern design” is not a measurable requirement. A hypothetical project could instead specify: “The user must find the invoice amount and status quickly; the accent color is reserved for the primary action; Persian text on mobile takes priority.” This is our suggestion for making a brief more precise, not a result from testing Anti-Slop.
The code skill also has a narrow scope. antislop-code says to change only comments, leaving executable logic, identifiers and code structure untouched. Removing comments that merely repeat the next line may make reading easier. It does not imply bug fixes or faster execution. Scope of antislop-code
The original installation instructions missed a step
According to the installation guide, the installer requires Node.js 20.12.0 or later:
npx antislop-aiThis starts an interactive process: you choose skills, installation scope and agent. A project installation can update the agent’s entry file with references to the skills. “One command” does not mean no choices or no file changes. Installer implementation
For Claude Code, the project’s instructions first add the marketplace, then install the plugin:
/plugin marketplace add https://github.com/miqdadbadjuber/anti-slop
/plugin install antislop@anti-slopThe installer defines routes for Claude Code, Codex, Hermes and OpenCode. This support is documented in the code and guide; we have not tested each agent ourselves. A new session is needed after installation. Installation routes
To assess its value for your work, compare a fixed brief under three conditions: no rules, Anti-Slop alone, and Anti-Slop with design direction. Keep the model and tools fixed and repeat each condition several times. Alongside appearance, record unsourced statistics, inactive buttons, mobile behavior and manual correction time. This is a proposed evaluation method; we have no measured improvement figure for the tool.
Source note: this review refers to version 3.2.25 at commit 941b153. The version and Node.js requirement are recorded in that snapshot.
Draft prepared with AI assistance; the reviewer’s name will be recorded after human review.