Detect-AI and Humanize: A Two-Sided Approach to AI Text Detection

Third in the series is a pair of skills from the humanizerai/agent-skills repo. Where blader’s humanizer (post one of this series) works from a list of stylistic tells, this project pairs a detection step with the rewrite step, and is explicit that its target is bypassing AI-detection tools like GPTZero, Turnitin, and Originality.ai.

Two commands, two jobs

Installing the repo gives you two separate skills: /detect-ai, which scores a piece of text from 0–100 for how likely it is to be flagged as AI-generated and returns a breakdown of the metrics behind that score, and /humanize, which does the actual rewrite to bring that score down.

That two-step structure is the interesting bit for me – it means you can check a document, see roughly where it’s getting flagged, rewrite it, and then re-check rather than rewriting blind and hoping.

Where this differs from humanizer

Both projects are aimed at similar output, but the framing is different: blader’s humanizer is styled as a prose-quality tool grounded in Wikipedia’s AI-writing guidelines, with a no-fabrication rule built in. This one is framed more directly around detector scores. Worth trying both on the same piece of text to see which rewrite you prefer.

Installing it

It installs as two separate skill directories rather than a single plugin:

git clone https://github.com/humanizerai/agent-skills.git
cp -r agent-skills/skills/detect-ai .claude/skills/
cp -r agent-skills/skills/humanize .claude/skills/

It also supports Cursor, Codex, and other agents that follow the same skills convention.

Next Sunday: Align, a skill that doesn’t just correct wrong claims from an LLM but tries to trace them back to whatever stale instruction or bad assumption caused them in the first place.