Editing AI-Sounding Writing: Which Skill Should You Use?

The useful question is not “which skill hides AI writing?” The useful question is “which skill helps a student turn a generic draft into a specific, accountable piece of writing?”

That distinction matters. Skills that market themselves as detector bypass tools are the wrong default for a learning wiki. They encourage the student to optimize against a classifier instead of improving the draft. The better workflow treats “AI taste” as an editing signal: vague claims, inflated transitions, symmetrical rhythm, fake confidence, missing human agency, and evidence-free polish.

This page is a decision guide. The longer Killing AI Slop entry covers stop-slop and avoid-ai-writing in more detail.

*Source: stop-slop on Skills stop-slop on GitHub avoid-ai-writing on Skills avoid-ai-writing on GitHub humanize on Skills remove-ai-style on Skills*

The decision map

Situation Use Why
You need a fast blunt pass before publishing stop-slop It scores directness, rhythm, trust, authenticity, and density. It is good at catching filler, false agency, adverbs, and polished-but-empty prose.
You need diagnosis, rewrite options, or file edits avoid-ai-writing It has detect, rewrite, and edit modes, plus context and voice profiles. It explains why a phrase was flagged instead of only banning it.
You want to “bypass detectors” Do not make that the default That framing trains the wrong habit. Use a writing-quality pass, then add evidence and author judgment.
The draft has no specific claims yet Use neither A skill cannot humanize an empty argument. Add names, dates, examples, numbers, observations, or a real position first.

Default recommendation

Use stop-slop first when the draft is short, public, and too smooth. It is deliberately strict. The core rules are easy to teach:

  • Cut filler phrases.
  • Name the human actor.
  • Replace vague abstractions with specific claims.
  • Vary sentence length.
  • Remove em dashes and formulaic contrasts.
  • Score the result out of 50.

This makes it a good classroom tool. A student can run the pass, see the score, then explain which edits improved the draft. The score is not the grade. It is a revision prompt.

Use avoid-ai-writing second when you need nuance. It can run in detect-only mode, rewrite mode, or edit a file in place. The context profiles matter: a technical blog should not be judged the same way as a LinkedIn post, and documentation should not be forced into a personal essay voice. The skill also has the right caution built into its premise: AI-writing signals are not proof of authorship.

A practical workflow

rough draft
  -> add real material       names, examples, dates, measurements, decisions
  -> stop-slop pass          score directness, rhythm, trust, authenticity, density
  -> avoid-ai-writing detect diagnose vocabulary, structure, and context mismatch
  -> revise manually         keep the author's voice and one or two natural irregularities
  -> final read              ask whether each paragraph adds something new

The order is intentional. If you run a “humanizer” before the draft contains evidence, the tool can only change surface style. It may sound better, but it still will not say anything.

For student work, I would ask for a short revision note with three rows:

Before After Why the edit improved the argument
Generic claim Specific claim Names the evidence or the actor
Smooth transition Direct bridge Shows how this paragraph follows the last
Detector-style polish Author voice Preserves the student’s actual judgment

That turns the skill from a camouflage tool into a writing lesson.

Install the useful defaults

The Skills CLI supports global, agent-specific installs. For a user-level setup that covers both Codex and Claude Code:

npx skills add hardikpandya/stop-slop --global --agent codex claude-code --skill stop-slop --yes
npx skills add conorbronsdon/avoid-ai-writing --global --agent codex claude-code --skill avoid-ai-writing --yes

After installation, the expected user-level folders are:

~/.codex/skills/stop-slop
~/.codex/skills/avoid-ai-writing
~/.claude/skills/stop-slop
~/.claude/skills/avoid-ai-writing

Then invoke them in plain language:

Use stop-slop on this draft. Score it and tell me the three edits that matter most.
Use avoid-ai-writing in detect mode on draft.md. Do not rewrite yet. Group issues by severity.
Use avoid-ai-writing to edit draft.md in place, technical-blog context, blunt voice. Preserve quotes and code blocks.

Candidates to test, not defaults

Some skill-list posts surface broader “humanizer” tools. Treat them as candidates, not the teaching default.

Candidate Status Why I would be careful
humanizerai/agent-skills@humanize Needs testing Its public skill description says it transforms text to bypass AI detectors and uses credits. That is the wrong primary frame for students.
zc277584121/marketing-skills@remove-ai-style Worth testing Its description includes deterministic Markdown analysis before rewrite, which is closer to an editorial workflow. Still verify on real drafts before recommending.
Generic AI detectors Weak signal only Detectors can help locate patterns, but they should not be used as proof that a student used AI.

Language-specific writing cleanup should live in separate posts with real examples. Mixing several languages into one skill list would make this page less useful.

What success looks like

A good pass does not make the prose invisible. It makes the writer more visible.

  • Each paragraph adds a claim, example, observation, or decision.
  • The draft names who did what instead of assigning agency to abstractions.
  • The rhythm varies because the idea demands it, not because a tool shuffled sentence lengths.
  • The student can explain what changed and why.
  • The final text would still sound like that student in office hours.

That is the line I would teach: remove the machine defaults, then restore the human responsibility.