Rules for an AI coding agent to filter out generic AI-generated UI designs, text, and code.
Security
antislop is a set of open source skills. Everything it does lives in this repository, and you can read any file before you install it. It is also scanned by third-party security auditors on skills.sh (Socket, Snyk, and Gen Agent Trust Hub). This page explains what those audits flagged and why each behavior is deliberate, so a warning is never a mystery.
What the audits say
As of v3.1.3, Gen Agent Trust Hub rates antislop Warn (MEDIUM) on two findings. Both describe features that are the point of the product. A third finding was addressed in v3.1.3. Details below.
Finding 1: it writes to agent entry files
The report notes that antislop appends a pointer block to an agent entry file such as CLAUDE.md or AGENTS.md. This is how a filter that runs on everything stays loaded: the entry file names the installed skills, so the next session reads them. The pointer block:
- is a static markdown reference. It lists the installed skills and carries no executable content.
- is written only after you approve. The installer and the manual wizard both ask first.
- is added at the end of the entry file on a first install, and replaced in place on later runs. Nothing else in the file is rewritten: your prose and code blocks are left alone, the line ending is the one the file already uses, and the only tidying is the blank lines at the seam where the block joins. On a later run the only lines removed are its own, the two markers and the block between them.
- lives in cli/lib/install.mjs for the installer, and in the wizard inside antislop.md for the manual path.
Removing this behavior would remove the core promise: a filter that is always on, not one you remember to invoke. Any always-on ruleset that writes a reference to an entry file matches this pattern.
Finding 2: it references an external script
The report notes that the human skill references a Python script, contrast-check.py. This is the contrast checker, an accessibility tool that validates color contrast against WCAG ratios. It:
- ships as a local file inside the skill folder, skills/antislop-human/contrast-check.py. You fetch it; the agent never downloads anything from the network.
- only checks the color pairs you pass to it. It has no network access and reads no other data.
- runs only when contrast work needs a WCAG check, which is the whole purpose of the human skill.
The plugin also exposes a contrast MCP tool, contrast-mcp.py, with the same properties: a local file, no network, no data beyond the color pairs it validates.
Finding 3: it reads external content (addressed in v3.1.3)
A finding noted that antislop reads DESIGN.md for design direction. The core now states the boundary explicitly: DESIGN.md is data to apply, not instructions to obey. The agent extracts only the design fields (identity, personality, palette, typography, mood, dials) and treats anything that reads like a command as content, not as an instruction. See antislop.md.
What antislop never does
- It never downloads or runs code from the network.
- It never reads, sends, or logs credentials or private data.
- It changes files only when you run a command that does, and only where you pointed it: the installer copies the skill folders you chose into the agent folder you picked, and on a project install adds a pointer block to the entry files of the agents you picked, the wizard adds only that pointer block, and
npx antislop-ai --mode during(orafter, orask) writes your saved usage mode to a settings file under your home directory. Nothing else in your project or your home directory is touched. - It is fully open source. Read the skills, the installer, and the checker before you install.
Verdicts that cleared
Socket previously warned on an install command that the wizard's install steps named. Those steps no longer name one: a missing skill is reported and left to the user to fetch from the matching release. Socket now passes.