Skip to content
Back to search

anti-ai-slop

SkillWorks with CodexGitHub

An agent skill with multiple gates to check and to remove any AI slop. Use it with codex/claude code or any of your favorite AI tools.

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/misbahsy--anti-ai-slop/badge.svg)](/listings/misbahsy--anti-ai-slop)

HTML

<a href="/listings/misbahsy--anti-ai-slop"><img src="/api/listings/misbahsy--anti-ai-slop/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

For Claude Code:

git clone https://github.com/misbahsy/anti-ai-slop.git
cp -r anti-ai-slop/skills/anti-ai-slop ~/.claude/skills/

For other agents, copy `skills/anti-ai-slop` into whatever folder your agent loads skills from.

The checks need Python 3.8 or newer. The optional model review also needs LiteLLM:

pip install litellm

The first time you ask for a model review, the agent asks where your API key is stored and runs everything for you. It saves where the key lives, never the key itself.

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

README.md · README · 179 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. # anti-ai-slop
  2. A skill for Claude Code, Codex, and other coding agents that removes AI-writing tells
  3. from a draft and then checks the result before handing it back.
  4. <p align="center">
  5. <img src="assets/gates.svg" alt="A draft is rewritten, then passes four gates: invisible characters, a linter, a second agent review, and an optional review by another model. Drafts that fail go back to the rewrite, up to two passes." width="100%">
  6. </p>
  7. ## What it catches
  8. ### Sentence patterns
  9. 1. **Negative parallelism.** "It isn't a pricing problem. It's a trust problem."
  10. 2. **Throat-clearing openers.** "Here's the thing," "Let me be honest"
  11. 3. **Faux-insight setups.** "What most teams get wrong," "The part nobody mentions"
  12. 4. **Rhetorical setups.** "Think about it:" "Sound familiar?"
  13. 5. **Colon reveals.** "The clever part: it retries on its own."
  14. 6. **Trailing -ing analysis.** "...adds dark mode, underscoring its focus on users"
  15. 7. **Importance puffery.** "stands as a testament to," "cannot be overstated"
  16. 8. **Telling the reader what to think.** "As you can see," "The key takeaway here is"
  17. 9. **Fake-strong verbs.** "serves as a unified layer," "has the ability to"
  18. 10. **False agency.** "The consensus emerged." "The data tells us"
  19. 11. **Dramatic fragments.** "That's it. That's the product."
  20. 12. **Fake-profound endings.** "And we're just getting started."
  21. 13. **Recap endings.** "In conclusion," "To sum up"
  22. 14. **Vague declaratives.** "The stakes are high." "The implications are significant."
  23. 15. **Stacked hedges.** "could potentially," "it might be argued that"
  24. 16. **Half-and-half framing.** "Pricing is half the story. The other half is..."
  25. 17. **Stiff wording.** "only when the cheaper option will not do"
  26. ### Habits that only show up in bulk
  27. 18. **Contrast on every claim.** "a settings change rather than a code change,"
  28. "instead of waiting." One is fine. Three in a short post gets flagged.
  29. 19. **Too many rhetorical questions** that the post answers itself
  30. ### Words and phrases
  31. 20. **AI vocabulary.** "delve," "leverage," "seamless," "tapestry," and about 120 more
  32. 21. **Empty phrases.** "at its core," "in today's fast-paced world," and about 100 more
  33. 22. **Canned openers.** "Moreover," "Ultimately," "In essence"
  34. 23. **Chatbot leftovers.** "Great question," "Hope this helps," "feel free to"
  35. 24. **Unsourced claims.** "studies show," "experts agree"
  36. 25. **Vague crowds.** "most companies," "we all know"
  37. 26. **Filler words.** "truly," "basically," "incredibly"
  38. 27. **Sweeping words.** "everyone," "nobody," "always"
  39. ### Rhythm and formatting
  40. 28. **Same-length sentences** four times in a row
  41. 29. **Short punchy sentences** stacked four deep
  42. 30. **Em dashes** everywhere
  43. 31. **Emoji used as headings or bullets**
  44. 32. **Lists of three** used for rhythm, over and over
  45. 33. **Adjective stacks.** "powerful and intuitive"
  46. 34. **Random bold** in the middle of sentences
  47. 35. **Headings over one or two sentences**
  48. 36. **Title Case Headings**
  49. 37. **Invisible characters** like zero-width spaces
  50. ### Things only a second reviewer catches
  51. 38. **Made-up facts.** A number, source, or quote the original never had
  52. 39. **Made-up connections.** A time frame, a cause, or a reaction added while smoothing
  53. a sentence, like "over several quarters" or "the feature customers ask about most"
  54. 40. **Synonym cycling.** "The agent reviews the draft. The assistant scores it. The
  55. tool suggests fixes."
  56. 41. **Changed meaning.** A dropped caveat, a claim made stronger than the original, or
  57. an argument that stops making sense after the cuts
  58. ## How it works
  59. You give your agent a draft. It rewrites it, and then the draft has to get through
  60. four gates before you see it:
  61. 1. **Gate 1: invisible characters.** Removes zero-width spaces, hidden tag characters,
  62. bidi controls, and odd spacing that come along with pasted AI text. Emoji and
  63. non-Latin text are left alone.
  64. 2. **Gate 2: the linter.** A Python script scores the draft out of 100 against
  65. everything in the list above that a script can find. That covers sentence templates,
  66. AI vocabulary, empty phrases, chatbot leftovers, unsourced claims, contrast habits,
  67. rhythm, and formatting. It skips code and quotes, gives the same score every time,
  68. and needs 90 with no serious findings.
  69. 3. **Gate 3: a second agent.** It gets the original and the rewrite, but not the first
  70. agent's reasoning. It catches what a script can't: made-up facts, made-up
  71. connections, changed meaning, dropped caveats, synonym cycling, and any slop the
  72. linter missed.
  73. 4. **Gate 4, optional: another model.** The same review, run by Gemini 3.8 Flash by
  74. default through [LiteLLM](https://github.com/BerriAI/litellm). You can swap in any
  75. provider or add several, and each one reports a human score and its cost.
  76. If a check fails, the draft goes back for another pass, up to two times. Anything
  77. still unresolved is listed for you instead of hidden.
  78. The second reviewer is there because the linter can't tell a real fact from an
  79. invented one. In testing, rewrites that scored 100 on the linter still had details the
  80. original never mentioned.
  81. By default the skill assumes a model wrote the draft and cuts hard. If you say you
  82. wrote it yourself, it keeps your voice and only fixes the obvious tells.
  83. ## Install
  84. For Claude Code:
  85. ```bash
  86. git clone https://github.com/misbahsy/anti-ai-slop.git
  87. cp -r anti-ai-slop/skills/anti-ai-slop ~/.claude/skills/
  88. ```
  89. For other agents, copy `skills/anti-ai-slop` into whatever folder your agent loads
  90. skills from.
  91. The checks need Python 3.8 or newer. The optional model review also needs LiteLLM:
  92. ```bash
  93. pip install litellm
  94. ```
  95. The first time you ask for a model review, the agent asks where your API key is stored
  96. and runs everything for you. It saves where the key lives, never the key itself.
  97. ## Use it
  98. Just ask your agent:
  99. - "De-slop this blog post before I publish it."
  100. - "ChatGPT wrote this email. Make it sound like me."
  101. - "Does this read as AI-written? Point out the patterns but don't rewrite it."
  102. - "Clean this up and have another model check it."
  103. You get back the edited draft plus a short list of what changed, including anything it
  104. cut because there was no source for it.
  105. ## Run the checks yourself
  106. ```bash
  107. python3 skills/anti-ai-slop/scripts/sanitize.py draft.md --fix
  108. python3 skills/anti-ai-slop/scripts/slopcheck.py draft.md
  109. python3 skills/anti-ai-slop/scripts/grade.py original.md rewrite.md
  110. ```
  111. `slopcheck.py` exits with 0 on a pass and 1 on a fail, so it works in a pre-commit
  112. hook or CI.
  113. If the linter flags something you want to keep, add a comment to that line:
  114. ```markdown
  115. Our leverage ratio fell to 2.1x. <!-- slop-ignore W1 -->
  116. ```
  117. ## Tests
  118. ```bash
  119. python3 skills/anti-ai-slop/tests/test_slopcheck.py
  120. python3 skills/anti-ai-slop/tests/test_credentials.py
  121. python3 skills/anti-ai-slop/tests/test_grade.py
  122. ```
  123. No network or API keys needed. One of the tests checks that the skill's own docs pass
  124. the linter.
  125. ## Adding rules
  126. Everything the linter looks for lives in `skills/anti-ai-slop/scripts/rules.json`. Add
  127. a word, phrase, or pattern there and rerun the tests. If the clean test file drops
  128. below 95, the new rule is flagging normal writing.
  129. ## Inspired by
  130. - [no-ai-slop](https://github.com/petergyang/no-ai-slop) by Peter Yang
  131. - [stop-slop](https://github.com/hardikpandya/stop-slop) by Hardik Pandya
  132. - [slop-cop](https://github.com/howshannon/slop-cop) by Shannon Tran
  133. - [clearmode](https://github.com/eugeniughelbur/clearmode) by Eugeniu Ghelbur
  134. ## License
  135. MIT

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.