HumanInk

Why it's ranked here

It turns a vague complaint, that a draft reads like a chatbot, into a checklist with named entries, before and after examples, and trigger phrases in several languages. The scoring is written out as an explicit formula with per-pattern weights, minimum counts, overlap discounts and a length adjustment, plus a worked example, so the logic is inspectable. MIT licensed and installable with a single clone. The limit is equally concrete: the number is computed by the model, not by code, so it is a structured opinion rather than a measurement.

What's good

The catalogue is practical. Each entry names the phrases to watch for and shows a real rewrite, such as turning a sentence that serves as a testament into one that just states the fact. Context modes change the weights sensibly: academic mode lowers the penalty on hedging and on words like Additionally, and creative mode relaxes the em dash rule. A confidence label drops to low when the result rests on one to three patterns, which is an honest hedge. Protected regions let you fence off quotes or code the rewrite must not touch.

Tradeoffs

Nothing in the repository measures whether the detector is right. There are no labelled samples, no accuracy figures and no tests, and the only automated check is a spelling pass over the Markdown. The claim that it can tell which vendor's model wrote a passage rests on listed pattern clusters, not on any published validation. The files also disagree with each other: the skill header says six languages while the contributor guide lists seven, and the maintainer notes point to a version field that the skill header does not carry.

How to use it well

Use it as an editing pass on drafts you already own: blog posts, reports, documentation, where the pattern report teaches you which habits to drop. The score only mode and the explain flag are the most useful for learning, since they show the reasoning without silently rewriting everything. Feed two or three samples of your own writing if you want output in your voice. Do not treat the number as evidence that someone else used AI; it is a model's reading of a checklist, and nothing here backs it with measured accuracy.

Observed

License

MIT

Implementation

Markdown instructions only, no executable code

Interfaces

Claude Code slash command; Claude.ai Project knowledge upload

Install

git clone into the Claude skills directory

Pattern catalogue

35 numbered patterns with severity weights in PATTERNS.md

Languages listed

EN, PT, ES, FR, DE, JA, IT per CONTRIBUTING.md; SKILL.md header says 6

Automated checks

Spell check workflow only; no tests or accuracy benchmark in files read