I got tired of rewriting AI articles. Humanizer gave me a better system
I got tired of rewriting AI articles. Humanizer gave me a better system
When I use AI to help with a long article, the first draft is rarely the real problem. The editing is.
I read the article, change a few paragraphs, remove words I would never use, and fix the rhythm. Then I read it again. Another section still sounds too polished. The headings feel mechanical. Every paragraph seems to end with a lesson. I edit it again.
Sometimes I repeat that cycle several times before the article sounds like something I would publish under my name.
That is why the open-source Humanizer project caught my attention. It is an agent skill for Codex, Claude Code, and other compatible coding assistants. Its job is simple: find common AI writing habits and rewrite the prose without changing the underlying claims.
The writer still makes the final call. The skill provides a more useful editing pass.
Why long AI drafts become tiring to edit
AI-generated writing can be grammatically clean while still feeling wrong.
The problem is often structural. The same patterns appear across the entire article:
- staged openings such as "Let's dive in"
- dramatic contrasts such as "It is not just a tool, it is a revolution"
- conclusions that repeat the previous paragraph
- forced lists of three
- broad claims dressed up as important insights
- decorative bold text and title-case headings
- vague authority such as "experts believe"
- chat leftovers such as "I hope this helps"
One occurrence may be harmless. A long article can contain dozens of them. Editing becomes repetitive because you are fixing the same habit in different forms.
Humanizer turns those recurring problems into a reusable editing process.
What the Humanizer skill does
Humanizer is written as a Markdown skill rather than a traditional application. Once installed, a compatible agent reads its instructions and applies them when reviewing text.
The repository currently documents 26 writing patterns. They cover staged language, repetitive rhythm, inflated claims, mechanical formatting, chat residue, and replies written for the wrong reader.
Its process has three useful parts:
- It identifies the strongest writing patterns first.
- It drafts a rewrite without treating the original paragraph structure as fixed.
- It checks whether the rewrite changed or invented any facts before producing the final version.
This check prevents a natural-sounding rewrite from changing a number, removing a condition, inventing a source, or turning uncertainty into certainty.
The project instructs the agent to preserve facts, meaning, code blocks, frontmatter, data, and link targets. When you point it at a file, it focuses on the prose.
What Humanizer does not promise
Humanizer is not marketed as a way to defeat AI detectors. Its own documentation says that detector evasion is not the goal and that detectors may still flag rewritten text.
Research has repeatedly found that AI-text detectors can be unreliable outside controlled conditions. A 2023 study by Sadasivan and colleagues showed that paraphrasing can reduce detection performance and that reliable detection becomes harder as generated text approaches human writing. Another peer-reviewed study found that several detectors incorrectly classified many essays written by non-native English writers as AI-generated.
A detector score should therefore not be your main editing target. A better target is whether the article is accurate, specific, useful, and recognizably yours.
If your goal is only to move a detector score, you may replace clear writing with awkward vocabulary and random sentence changes. The article can become less useful while appearing more "human" to one classifier.
How to install Humanizer
Codex
Run this command in your terminal:
npx skills add blader/humanizer --global --agent codex
Remove --global if you only want the skill inside the current project.
After installation, restart or reload your agent session so it can discover the new skill.
Claude Code
The repository provides a plugin installation route:
/plugin marketplace add blader/humanizer
/plugin install humanizer@humanizer
You can then call:
/humanizer:humanizer
The repository also provides a Skills CLI option for Claude Code:
npx skills add blader/humanizer --global --agent claude-code
Manual installation
Because the skill is Markdown, you can download the repository and copy SKILL.md into the skill folder used by your agent. Check your agent's documentation for the correct directory.
My workflow for a long article
Installing the skill is the easy part. The quality comes from when and how you use it.
1. Build the factual draft first
Start with the argument, examples, data, and sources. Do not ask the tool to humanize a draft that still has missing evidence or an unclear point.
Humanizer is an editor. It does not verify your research automatically.
For a factual article, I would finish these items first:
- the main claim
- supporting evidence
- source links
- personal experience
- practical instructions
- any limits or uncertainty
2. Give it a sample of your real writing
This is the most useful step if you publish regularly.
Paste two or three paragraphs you wrote yourself. Choose a sample that sounds like the voice you want for the article. Then ask the skill to match its sentence length, vocabulary, punctuation, and rhythm.
Use a prompt like this:
/humanizer
Here is a sample of my natural writing:
[PASTE TWO OR THREE PARAGRAPHS]
Now humanize the article below. Preserve every factual claim, number,
quotation, citation, heading link, and technical instruction. Remove repeated
AI writing patterns, but do not make the article more dramatic or promotional.
[PASTE ARTICLE]
Without a writing sample, the tool can remove obvious patterns. With a sample, it has a better chance of sounding like you.
3. Use file mode for Markdown articles
If the article already exists in your project, point the agent to the file:
Humanize the prose in content/blog/my-article.md.
Preserve the frontmatter, code blocks, tables, URLs, citations, and factual
claims. Do not change the slug or metadata values.
This is safer than copying a long article into the chat, especially when the file contains code or structured metadata.
Commit the original file or save a copy before the rewrite. You should always be able to compare the changes or restore a sentence the tool removed.
4. Review the diff, not only the final article
Read what changed.
Check whether the rewrite:
- removed a qualification that mattered
- changed a technical meaning
- flattened a personal detail
- altered a quotation
- weakened a source attribution
- made every paragraph sound equally casual
Human writing is uneven. Some sentences are short. Others need room. A tool can remove obvious AI habits and still smooth away the parts that make the article yours.
5. Read the article aloud once
Reading aloud catches problems that a detector cannot measure. You hear repeated openings, unnatural pauses, oversized sentences, and phrases you would never say.
If you stumble on a sentence, simplify it. If a paragraph has no new information, cut it. If the conclusion repeats the introduction, end earlier.
One careful read is usually more useful than sending the same article through five different rewriting tools.
A practical example
An AI draft might say:
In today's rapidly evolving digital landscape, authentic communication is more important than ever. Humanizer is not just a writing tool, it is a revolutionary solution that empowers creators to unlock the full potential of their unique voice.
A cleaner version would be:
Humanizer removes writing patterns that often make AI-assisted drafts sound generic. It keeps the meaning while giving the writer a shorter, clearer draft to review.
The second version states what the tool does and stops. It does not need unusual vocabulary to sound natural.
Where this fits in a business workflow
For a founder, consultant, or technical team, Humanizer can become the final prose review before publication.
A simple workflow could be:
- Research and collect sources.
- Draft the article with AI assistance.
- Verify every factual claim and link.
- Run Humanizer with a real writing sample.
- Review the diff manually.
- Read the final copy aloud.
- Publish only after a person approves it.
This can also be automated. An agent can watch a drafts folder, create a humanized review copy, preserve the original, and produce a change summary. A person still approves the final version.
The workflow saves time because the repeated cleanup happens consistently. The human review stays focused on judgment, accuracy, and voice.
When not to use it
Do not run a humanizing pass over legal wording, compliance statements, contracts, medical instructions, direct quotations, or regulated disclosures without strict review. Small stylistic changes can alter meaning.
It is also unnecessary for every message. A short email may need one direct edit, not a full rewriting workflow.
Most importantly, do not use it to hide fabricated evidence, copied work, or undisclosed automation where disclosure is required. Better phrasing does not fix unreliable content.
The lesson from my own editing process
I used to treat every article as a new cleanup job. I would edit the same habits again and again because I had no repeatable standard for what "sounds like AI" meant.
Humanizer makes those habits visible. That gives me a stronger starting point, but I still own the final article. I choose the sources, keep the details that sound like me, and decide what deserves to stay.
My finish line is an article I can read later and still recognize as mine. A detector cannot make that judgment for me.
If you want more practical AI workflows, follow me on Instagram @salama_win and connect with me on LinkedIn.
If your business needs an AI system, automation workflow, or production implementation, email me and tell me what process you want to improve.
Sources
Stay ahead of the curve
Join my private newsletter for exclusive insights, tools, and thoughts straight to your inbox. No spam, just value.