smartwatermelon/personify

An agent skill that keeps text from reading as machine-written before anyone else reads it: emails, PR descriptions and review comments, docs, blog drafts, essays. The draft follows your own register, and then a detector checks the result rather than the skill declaring it clean.
Personify 2.0 is a break from 1.x. The hand-maintained taxonomy of AI-writing tells is gone, along with the A/B harness and the evidence corpus that fed it. Three things replace them: a per-user voice guide, current model judgment about what reads as machine-written, and Pangram, which returns a verdict on the finished text.
It started as a structural cousin of blader/humanizer, with the same idea and the same MIT license, and LICENSE credits it. The wording, the rules, and the version history here are independent and were never synced against that project.
## How it works
1. Draft, guided by your VOICE.md and seven hard rules, among them no dashes, never invent a fact, preserve every fact, and name the actor.
2. Check. The result goes to scripts/pangramcheck.py, which submits it to Pangram once and reports a verdict.
3. A Human verdict passes. Anything else stops and reports the score and the flagged spans.
There is no edit loop, and this is deliberate. Editing a draft toward the detector was measured and does not work: two rounds of rewriting moved the score from 0.990 to 0.989 to 1.000, so the second rewrite scored worse than the original. A failing verdict means rewriting from the source material, or routing the text to a human. It never means editing harder.
Text under 40 words is skipped rather than classified. Below that the detector produces false passes, including the dangerous direction where AI-written prose comes back Human.
## What a detector cannot see
Two rules survive independently of Pangram, because Pangram scores how prose reads and cannot see how an artifact is shaped. It will pass a pull request description carrying ceremonial headers, bolded labels, and a "ready to merge upon approval" sign-off, as long as the sentences read human.
GitHub PR descriptions. No headers, no bullets, no bold, no dashes. Four parts as plain prose: the problem, the evidence, the solution, references.
Code comments. Never more than 1:1 comment lines to code lines, and far lower in practice. A comment is for what the code cannot say.
Both are universal rules and both outrank a personal VOICE.md, which is otherwise authoritative. They also live in rules/structure.md for a caller that wants them without the rest of the skill.
## Installation
The runtime artifact is the skill directory, not SKILL.md alone: scripts/pangramcheck.py, scripts/lengthcheck.py, rules/structure.md, and VOICE.example.md ship with it. Step 0 falls back to VOICE.example.md from the skill directory when no voice guide exists.
### Claude Code plugin
/plugin marketplace add twistedmelonman/personify
/plugin install personify@personify
Once installed, invoke it as /personify:personify.
### Claude Code, project-local or global
Copy the skill directory so the script comes with it:
bash
mkdir -p .claude/skills/personify # or ~/.claude/skills/personify