The problem I was actually solving
The draft did not need more personality for its own sake. It needed less template.
The symptoms were familiar: a title that sounded like a category label, section headings that could have come from any product comparison, too many tidy lists, and a conclusion that was correct but bloodless. The page was useful. It just sounded like it had been assembled.
That matters for AI Picker because the site is not trying to be a giant tool directory. It is trying to be a small collection of judgment notes. If the writing starts to sound like a generic agent report, the trust signal drops.
What I did not want
I ruled out one path early: chasing AI detector scores. That is the wrong incentive for this site.
- I did not want a tool that rewrites everything just to sound more casual.
- I did not want fake human texture: random asides, forced slang, or deliberate messiness.
- I did not want to lose official-source caveats, pricing caution, or traceable links.
- I did not want a workflow whose main promise was bypassing GPTZero, Turnitin, or similar detectors.
The target was narrower: find the places where the draft stopped sounding like a person making a judgment.
The main candidates I found
This was my shortlist. The star counts are a rough heat signal from GitHub when I checked, not a durable quality score.
| Skill | Heat when checked | What it does | Usefulness for AI Picker |
|---|---|---|---|
| blader/humanizer | ~23.2k GitHub stars | Claude Code / OpenCode skill for removing signs of AI-generated writing; includes voice calibration and a second audit pass. | Very visible and worth studying, but too easy to use as a general-purpose “make it human” button. |
| conorbronsdon/avoid-ai-writing | ~1.8k GitHub stars | Portable skill for Claude Code, OpenClaw, Hermes, and agent-skills-compatible agents; supports rewrite, detect-only, edit-in-place, and voice profiles. | Best fit. It works like an editing gate: detect first, then make small changes. |
| AgriciDaniel/claude-blog | ~1k GitHub stars | A full blog production skill suite with sub-skills, agents, SEO checks, citation workflows, and a delivery contract. | Impressive, but too heavy for AI Picker. It solves scale. I needed editorial restraint. |
| brandonwise/humanizer | ~89 GitHub stars | OpenClaw skill and standalone CLI that scans for AI writing patterns, vocabulary tells, burstiness, type-token ratio, and readability metrics. | Good OpenClaw fit and useful as a diagnostic idea, but less proven than the larger options. |
| Aboudjem/humanizer-skill | ~74 GitHub stars | Pure Markdown humanizer with 43 AI patterns, 5 voice profiles, a 0-100 AI-tell score, and zero dependencies. | Useful checklist material. I would borrow the pattern list before using the score as a target. |
| jpeggdev/humanize-writing | ~25 GitHub stars | Claude Code skill with an 8-pass editing process covering structure, inflated language, AI vocabulary, rhythm, vague attribution, and connective tissue. | Good editing vocabulary, but early. Better as inspiration than the main gate. |
| writing-prose-like-a-human-for-agents | ~9 GitHub stars | A skill and subagent built around five plain rules: cut significance inflation, use plain verbs, end sentences at the fact, vary rhythm, and earn every adjective. | Low heat, high signal. The principles match the AI Picker voice better than most “humanizer” promises. |
Why I chose avoid-ai-writing
The deciding point was not popularity. If popularity were the only filter, I would have picked blader/humanizer. The deciding point was control.
AI Picker needs a writing cleanup step that can say: here are the AI-isms, here is what is probably a real problem, here is what might be intentional, and here are the minimal edits. That is closer to how avoid-ai-writing is positioned.
Its detect-only mode matters. Sometimes a table is useful. Sometimes a checklist is the right format. Sometimes a phrase sounds formal because the topic is formal. A good editing gate should not punish every structured sentence.
Edit-in-place also matters. I do not want a full rewrite that quietly changes a caveat, breaks a claim, or removes a source link. I want a small pass after the facts are already checked.
What I borrowed from the others
I still took notes from the smaller projects.
- From writing-prose-like-a-human-for-agents: end sentences at the fact, use plain verbs, and earn every adjective.
- From humanize-writing: watch for tidy section structures, repeated list lengths, connective tissue, and rhythm.
- From humanizer-skill: pattern lists can be useful, but scores should not become the goal.
- From blader/humanizer: voice calibration is valuable, but only if it preserves the author’s judgment instead of smoothing it out.
What I ruled out
I would not make a detector-bypass tool part of the AI Picker editorial workflow. For example, HumanizerAI Agent Skills explicitly markets detection and humanization around bypassing detectors. That may be what some users want. It is not what this site needs.
The reason is simple: a detector score is not the same as trust. A page can pass a detector and still be shallow, unverified, or misleading. AI Picker should optimize for traceable judgment.
How this fits the AI Picker workflow
The editing gate now sits near the end of the Research Notes process:
- Write the draft around a real decision question.
- Check official sources before making factual claims.
- Keep pricing claims conservative unless the source is verified.
- Use market voice only as signal, not proof.
- Run the writing cleanup skill in detect or audit mode.
- Make targeted edits, then reread the page for factual drift.
That last step is important. A writing cleanup pass can improve voice and still damage accuracy. The editor has to check both.
What changed in practice
The first page I used it on was the note about Agent-Reach and last30days. The original title was:
Nothing was wrong with it. But it sounded like a taxonomy. The revised title became:
That is a small change, but it moves the page from classification toward the reader’s actual question.
The same thing happened inside the article. “The buyer problem” became “I would not compare them by feature count.” “The short distinction” became “The split I would use.” The facts stayed. The voice became less assembled.
My bottom line
The useful writing cleanup skill is not the one that promises to beat detectors. It is the one that helps you notice where the draft stopped sounding like a person making a judgment.
For AI Picker, that means avoid-ai-writing is the main tool, and the other skills are reference material. The goal is not fake human copy. The goal is clearer judgment with fewer template sentences.