What is AI slop? Generic writing and code from agents
By Flavio
AI slop is generic writing or code a model repeats until drafts sound the same. Claude Code, Codex, Cursor and Grok Build ship no filter for it.
AI slop is writing or code that came from a model and still carries the habits that model repeats. The facts can be right. The draft still reads like every other draft: a bloated opening, a forced list of three, an abstraction with one caller, a comment that says what the next line already says.
Choosing anti-slop skills is the post that owns the review pass. This entry is the short definition. Slop grenades is the separate case, when that draft gets sent to someone else before you read it.
How AI slop shows up
The first pass from a coding agent is often useful and still generic. The habits are the same across tasks because they sit in the model’s defaults, not in your repo.
An anti-slop skill is a second pass. It is a SKILL.md file of review instructions. The agent scans what it just wrote and rewrites the generic parts. You then read the diff, because the skill can remove a sentence you wanted.
Slopsquatting is a different failure. That one is a package name the model invented, which someone else can register. Slop here is the prose and the code you already have in the file.
AI slop in each tool
None of the four product indexes uses the word slop, and none documents a mode that strips generic output for you. The cleanup that this site’s posts describe is a skill you add.
| Claude Code | Codex | Cursor | Grok Build | |
|---|---|---|---|---|
| A built-in slop filter | Not documented | Not documented | Not documented | Not documented |
| How a skill runs | The Skill tool, or a slash command the skill defines | $ plus the name, in the unslop post | / plus the skill name, in the unslop post | A user-invocable skill becomes /<skill-name>. The page does not name unslop |
Claude Code
The tools reference says a skill runs through the existing Skill tool. It does not add a new tool called slop. How to use the unslop skill shows /unslop in Claude Code for the skill in Cursor’s plugins repo. That command is the skill’s, not a built-in Claude Code command. See the tools reference.
OpenAI Codex
The skills page is where reusable instructions live. The agent guide does not mention a review pass for generic writing. The unslop post shows $unslop in Codex for that same skill file. See Skills in Codex.
Cursor
Cursor’s agent page lists tools for search, edits, and the shell. It does not list a slop filter. The unslop post shows /unslop in Cursor, and it says the skill ships inside pstack in Cursor’s plugins repository. See Cursor Agent.
Grok Build
Skills are folders Grok discovers, including .grok/skills/ and ~/.grok/skills/. A user-invocable skill also shows up as /<skill-name>. user-invocable defaults to true. The page does not name an unslop skill or a slop mode.
Grok also reads Claude Code skills, with no extra setup, alongside the .grok/ folders. A skill Claude Code already has can be visible there. The page still does not name unslop. See Skills, plugins, and marketplaces.
Details checked against each vendor’s documentation in November 2026.
A real example
The unslop post’s job is two steps: scan for the patterns, then rewrite while keeping the meaning and the tone. A prompt in that shape, after the draft exists:
Run the unslop skill on the paragraph I pasted below. Keep the meaning and the tone. Do not add a source I did not give you.
Skill Cabinet is the Mac app that shows which agents load a skill folder. The skills-folder tutorial is the map of those folders, so you can see whether the anti-slop skill is on the agent you are about to run.
When to use an anti-slop pass
The anti-slop post runs it after the draft, as its own pass, not while the agent is still deciding the structure. Use it when the writing or the patch is done and you want the repeated habits pulled out.
Skip it when the problem is a missing fact or a wrong command. The same post says a skill cannot supply a fact you never had, and it cannot make untested code work.
Slop grenades is the reason to read the result before you send the email or open the pull request. The pass does not replace that read.
Pitfalls
The anti-slop post’s limit is the one to keep. The skill can delete a sentence you liked, flatten an unusual phrase, or swap one repeated habit for another. You still read the diff.
A public skill is a starting list of patterns for one review pass. Project rules, such as how a confirmation must work, stay in the project’s own instructions. The anti-slop post keeps those out of the skill on purpose.
/unslop and $unslop only do something after that skill is installed. The unslop post is the install for Cursor, Claude Code, and Codex. On Grok, the skills page turns a user-invocable skill into /<skill-name> once Grok has discovered the folder, including through the Claude Code skills it reads. The page does not ship that skill for you.
Related on flaviocopes.com
- Choosing anti-slop skills ranks the skills and separates a writing pass from a code pass.
- How to use the unslop skill is the
SKILL.mdfile, the 28 rules, and the commands in Cursor, Claude Code, and Codex. - Slop grenades is what happens when the unread draft is sent to someone else.
- Skill Cabinet shows which agents load each skill. The folder tutorial lists the directories.
Related glossary entries
- Agent skills: An agent skill is a SKILL.md folder that teaches a coding agent one task. Where Claude Code, Codex, Cursor and Grok Build load it, and when to skip one.
- Slopsquatting: Slopsquatting is registering a package name a language model invented, so a later install of that name runs someone else's code. What a 2025 study measured.
Want me to talk about your product? You can sponsor this site.