AI Humanizer Skill for Claude and Codex, in Your Voice

If you already use Claude Code or Codex CLI, you probably do not need another AI humanizer subscription. You need a better workflow for the agent you already have.

To humanize AI text in your own voice, install the Agentic Humanizer skill, give Claude or Codex a real writing sample, and let the skill run a five-pass rewrite loop. Slop or Not is optional: it adds local detector scoring, readability, and Text Cleanup when you want measurement.

Most humanizer sites give you a text box and a score. That can be useful, but it is also narrow. Claude and Codex already understand instructions, can inspect source files, can ask follow-up questions, and can repeat a rewrite with a different strategy. The open-source Slop or Not plugin turns that into a repeatable tutorial workflow.

This guide covers setup, voice samples, the loop inside the skill, and the optional Slop or Not checks. It does not promise that any rewrite will pass every AI detector.

How Do You Humanize AI Text With Claude or Codex?
What Should You Install First?
How Do You Make the Rewrite Sound Like You?
What Happens Inside the Agentic Humanizer Skill?
How Do You Add Slop or Not Checks?
Can This Replace Paid AI Humanizer Subscriptions?
FAQ

How Do You Humanize AI Text With Claude or Codex?

Humanize AI text with Claude or Codex by treating the model as the editor, not by sending the draft to a separate paraphrasing site. Give the agent a writing sample, set tone and reading-level constraints, then run the Agentic Humanizer skill so each pass changes a different part of the draft.

The bad prompt is short:

make this sound human

That usually swaps one generic voice for another. The better version is a workflow:

Start with the AI draft you want to fix.
Give Claude or Codex a sample of writing that actually sounds like you.
Choose dialect, reading level, tone, and length.
Let the skill run targeted rewrite passes instead of one big paraphrase.
Read the final text yourself before you use it.

That is why Claude and Codex are the best humanizer option for many people. The model is already there. You do not need to pay a separate humanizer site just to make the same model follow a better process.

The Agentic Humanizer skill supplies that process. It is a set of instructions for the agent, not a hosted rewrite service. Claude or Codex does the writing work. Slop or Not can measure the result later if you want local AI detector feedback.

What Should You Install First?

Install the Slop or Not plugin bundle first when your agent supports plugins. It includes the Agentic Humanizer skill for rewriting and the Slop Check skill for direct local checks. The core humanizer works without the Slop or Not app; install the Mac app only when you want on-device detector scoring.

For Codex CLI, add the marketplace:

codex plugin marketplace add numen-tech/slopornot

Then run codex, open /plugins, switch to the slopornot marketplace, and choose Install plugin.

For Claude Code, run these commands inside Claude Code:

/plugin marketplace add numen-tech/slopornot
/plugin install slopornot@slopornot

Claude Code plugin installs namespace the command:

/slopornot:agentic-humanizer

Codex and most direct skill installs use:

/agentic-humanizer

For a non-interactive Claude Code install, first add the marketplace from an interactive session, then run:

claude plugin install slopornot@slopornot

For Claude Code, Cursor, or Windsurf without the plugin marketplace:

npx skills add numen-tech/slopornot

For Codex CLI, Gemini CLI, or OpenCode direct installs, clone the repository once and copy the skill directories into your harness:

git clone https://github.com/numen-tech/slopornot /tmp/slopornot
 
# Codex CLI
rm -rf ~/.codex/skills/agentic-humanizer ~/.codex/skills/slop-check
mkdir -p ~/.codex/skills
cp -R /tmp/slopornot/skills/agentic-humanizer ~/.codex/skills/
cp -R /tmp/slopornot/skills/slop-check ~/.codex/skills/
 
# Gemini CLI
rm -rf ~/.gemini/skills/agentic-humanizer ~/.gemini/skills/slop-check
mkdir -p ~/.gemini/skills
cp -R /tmp/slopornot/skills/agentic-humanizer ~/.gemini/skills/
cp -R /tmp/slopornot/skills/slop-check ~/.gemini/skills/
 
# OpenCode
rm -rf ~/.config/opencode/skills/agentic-humanizer ~/.config/opencode/skills/slop-check
mkdir -p ~/.config/opencode/skills
cp -R /tmp/slopornot/skills/agentic-humanizer ~/.config/opencode/skills/
cp -R /tmp/slopornot/skills/slop-check ~/.config/opencode/skills/

Restart the agent after a direct install so it discovers the new skills.

How Do You Make the Rewrite Sound Like You?

Make the rewrite sound like you by letting the skill ask its first-run questions before it edits. Agentic Humanizer asks about dialect, reading level, tone, length, saved defaults, and whether you want it to mimic a writing sample. That interview is the prompt most people should use.

Claude Code Agentic Humanizer prompt for dialect, reading level, tone, and length settings — Claude Code reviews the first-run Agentic Humanizer settings before the rewrite starts.

Start the skill and paste your draft:

/agentic-humanizer
[paste the AI draft]

Claude Code plugin installs use:

/slopornot:agentic-humanizer
[paste the AI draft]

Claude Code Agentic Humanizer prompt asking whether to mimic a writing sample for voice matching — Voice matching is optional, but it is the part that makes the rewrite sound less generic.

When the skill asks whether to mimic a writing sample, say yes and paste writing that already sounds like you. A few paragraphs are enough for a first try. Use 200 or more words when you can; 50 words is the hard minimum.

The skill then extracts style traits such as rhythm, register, contractions, punctuation, paragraph shape, and phrases you tend to avoid. It asks you to approve that read before the rewrite uses it. If the read feels wrong, correct it or rerun the extraction.

If you expect to reuse the same dialect, tone, and reading level, save the profile. That keeps future humanizer runs shorter without locking you out of changing the settings later.

Saved profiles keep repeated Claude Code humanizer runs from asking the same setup questions.

Keep the sample boring and safe: no client names, student records, medical notes, legal details, passwords, or private account numbers. Slop or Not checks run locally when you use the Mac app, but the writing sample is handled by Claude, Codex, or whichever agent runs the skill. The skill source has the advanced voice=/path, voice=off, and saved-profile details.

What Happens Inside the Agentic Humanizer Skill?

The Agentic Humanizer skill is an orchestrated rewrite loop. It resolves your preferences, optionally extracts a voice fingerprint, probes for Slop or Not Pro, then runs five different rewrite strategies. Each pass changes a different axis so the agent does not keep paraphrasing the same sentence.

The source describes seven orchestration steps:

Detect the host harness, such as Claude Code, Codex CLI, Cursor, Gemini CLI, OpenCode, or a generic client.
Handle saved profile and voice-management commands.
Resolve dialect, reading level, tone, and length.
Resolve a voice sample and cache a stylometric fingerprint.
Probe whether Slop or Not Pro is reachable through MCP or the CLI.
Run the five-pass humanization loop.
Return the final text, loop history, highest-impact edits, and cleanup stats when real Slop or Not cleanup ran.

The five rewrite passes are the useful part:

Pass	What the Agent Attacks
1. Pattern surgery	The most visible AI-writing tells, such as filler transitions, forced balance, vague claims, and repeated sentence shapes.
2. Dialect and tone	American or British spelling, tone, register, contractions, hedges, and voice-fingerprint fields.
3. Grade gap	Sentence length and word complexity until the draft sits near the target reading level.
4. Clean and targeted	Residual mechanical artifacts, hidden characters, odd punctuation, and the 1-2 tells still showing.
5. Structural pass	Sentence openings, paragraph rhythm, over-neat rule-of-three phrasing, and concrete examples supported by the source.

Voice matching does not touch every pass. The fingerprint is used in pass 2 for register, contractions, hedges, and function words. It returns in pass 5 for sentence openings, idioms, and paragraph rhythm. That keeps the sample from overpowering the task.

In core mode, the skill runs without detector calls and reports score and grade as n/a. With Slop or Not Pro active, it can run Text Cleanup before the baseline, score each pass locally, check Flesch-Kincaid grade level, clean the selected final text, and show cleanup counts.

How Do You Add Slop or Not Checks?

Add Slop or Not checks when you want Agentic Humanizer to run as a measured bypass loop, not a blind rewrite. Claude or Codex rewrites the text. Slop or Not scores each pass locally, checks readability, cleans hidden artifacts, and gives the agent feedback for the next pass.

That design gives you the best chance of lowering AI detector scores while keeping the rewrite in your own voice. Instead of asking "does this sound human?" once, the agent starts with a baseline, rewrites against specific AI tells, scores the new version, checks whether the reading level still fits, then changes strategy if the score is still too high.

With Slop or Not Pro active, the loop can:

Run Text Cleanup before the baseline so invisible characters, homoglyphs, and odd punctuation do not skew the draft.
Score the original text with the on-device AI text detector.
Check Flesch-Kincaid grade level so the rewrite does not become unnaturally simple or stiff.
Run each Agentic Humanizer pass with detector feedback from the previous pass.
Pick the best measured result, then clean the final text again before returning it.

No AI humanizer can promise a universal bypass across GPTZero, Turnitin, Originality.ai, and every future detector. The point is more practical: Slop or Not gives Claude or Codex a local scorecard so the humanizer can adapt instead of guessing.

To enable that measured loop, install Slop or Not for Mac, open it once, and activate Pro from Settings > Subscription. The CLI and MCP server are Pro features bundled inside the Mac app:

/Applications/Slop Or Not.app/Contents/MacOS/slop

Check status:

"/Applications/Slop Or Not.app/Contents/MacOS/slop" status --json

For Claude Code, register the local Model Context Protocol server:

claude mcp add --transport stdio --scope user SlopOrNot -- "/Applications/Slop Or Not.app/Contents/MacOS/slop" mcp

For Codex CLI:

codex mcp add SlopOrNot -- "/Applications/Slop Or Not.app/Contents/MacOS/slop" mcp

Or add this to ~/.codex/config.toml:

[mcp_servers.SlopOrNot]
command = "/Applications/Slop Or Not.app/Contents/MacOS/slop"
args = ["mcp"]

After restart, ask your agent to run slop_status. If that works, Agentic Humanizer can call detect_text, analyze_readability, and clean_text during the loop. The AI Detector MCP guide covers the setup in more detail, and the CLI page is better when you want explicit Terminal commands.

Can This Replace Paid AI Humanizer Subscriptions?

Yes, for many users. If you already pay for Claude or use Codex, an expensive AI humanizer subscription often gives you a thinner version of the model you already have. Most humanizer tools are wrappers: paste text, click a button, get one rewrite, then trust the score.

Claude and Codex are better humanizers because they can work with context. They can ask what tone you want, read your writing sample, preserve citations, respect project files, keep product claims accurate, and revise again when the first pass misses. A wrapper usually sees one box of text. An agent can see the job.

Agentic Humanizer adds the structure those models need. It asks first-run questions, matches your voice when you opt in, runs five targeted passes, and returns the loop history so you can see what changed. With Slop or Not Pro active, the same loop gets local detector scores, readability checks, and Text Cleanup feedback between passes.

That is the real subscription argument. You do not need to rent another humanizer if Claude or Codex can already do the rewriting and the Slop or Not skill can tell it how to do the work. Spend the effort on a good writing sample and a measured loop, not another generic rewrite box.

Dedicated humanizer sites can still be convenient when you need a browser-only workflow or a tool your school or team already approved. But if the goal is to humanize AI text in your own words, Claude or Codex with Agentic Humanizer is the better fit. For detector-focused workflows, read the full agentic AI Humanizer guide. For readability targets, see how Slop or Not readability scores work.

FAQ

Is Claude or Codex really the best AI humanizer?

For many people, yes. Claude and Codex are already strong editors, and the Agentic Humanizer skill gives them a repeatable workflow: preferences, voice matching, five targeted passes, and review notes. A separate humanizer subscription is less useful when your existing agent can do the rewrite with more context.

Do I need Slop or Not to humanize AI text?

No. The Agentic Humanizer skill works without Slop or Not. Claude or Codex can run the core rewrite loop and match a writing sample by itself. Slop or Not adds optional local AI detector scoring, readability checks, Text Cleanup, and cleanup stats on Mac with Pro active.

Does Slop or Not upload my writing sample?

No. Slop or Not does not upload the text it checks. The writing sample is handled by the agent doing the rewrite, not by Slop or Not. If Claude, Codex, or another hosted agent extracts the voice fingerprint, that sample follows the agent's own privacy model.

How many words should my voice sample have?

Use 200 or more words when you can. The skill has a hard minimum of 50 words and uses the first 3000 words for extraction. Pick writing that matches the target genre: an old essay for an essay, old emails for email, or a published post for blog copy.

Can this guarantee that AI detectors will say the text is human?

No. AI detectors are probabilistic, and different detectors disagree. The honest goal is to remove generic AI-writing patterns, make the draft sound like you, and check the result before use. A low score from one detector is a signal, not a universal guarantee.

Is this only for Claude Code?

No. The plugin supports Claude Code, Codex CLI, Cursor, Gemini CLI, OpenCode, and other agent tools that can load skills. Claude Code uses /slopornot:agentic-humanizer for plugin installs. Codex and most direct installs use /agentic-humanizer.

Try the Workflow

Install the Agentic Humanizer skill, save a voice sample at ~/.agentic-humanizer/voice.txt, and run the skill in Claude or Codex with the draft you want to fix. Add Slop or Not for Mac only when you want local AI detector scores, readability, Text Cleanup, and cleanup stats inside the loop.