jev-humanizer

Agent skill where Jev flags 35 AI-writing patterns per paragraph and the model edits only the flagged spots.

ProjectbetaMIT

Overview

jev-humanizer is an anti-slop skill for AI agents such as Claude Code and Codex, documented in Russian and based on smixs/humanizer-ru and blader/humanizer. It removes bureaucratic phrasing, cliches, false contrasts, long dashes and filler that mark text as AI-written while keeping facts and figures. Instead of asking the model to find its own slop, Jev checks every paragraph against 35 semantic patterns, a linter adds mechanical checks such as dashes and arrows, and the model receives a table of findings and edits only the marked places.

The author argues a model does not notice its own favourite turns of phrase while a classifier does, at under a cent per post. The reference and linter currently cover Russian only, though Jev reads English: on an English example it caught 'not just X, it's Y', a long dash and emoji but missed 'Let that sink in' and a rhetorical question. An audit mode returns a table of evidence and a verdict without editing. It needs Python 3 and a Jev key.

What the decision model decides

Type	Question	Answers
Yes / no	Does this paragraph show each of 35 AI-writing patterns?	—

How it works

Agent drafts or receives text
Jev checks each paragraph for 35 patterns; linter adds mechanical checks
Model gets a findings table and edits only flagged spots
Or run an audit-only verdict

Strengths

Targeted edits rather than full rewrites
Cheap per post
Audit mode without changes

Caveats

Pattern reference and linter are Russian-only
Misses some English patterns

Getting started

Clone to ~/projects/jev-humanizer, symlink skill/ into ~/.claude/skills or ~/.codex/skills, and put TYPESAFE_API_KEY in ~/.config/typesafe.env.

At a glance

Type: Project
Status: beta
Language: Python
Runtime: Agent skill and CLI
Has code: yes
License: MIT

Use cases and integrations

Verification & Guardrails272

How curated lists describe it

This project integrates Jev to provide structured decisions for its workflow. See the repository for implementation details.
logicrw/awesome-jev-projects
Антислоп скилл для AI-агентов: Jev от TypeSafe находит признаки нейросети в каждом абзаце, модель правит только найденное. На основе smixs/humanizer-ru и blader/humanizer.
hellogumbo/awesome-jev

Frequently asked questions

What does Jev decide?

Which of 35 AI-writing patterns appear in each paragraph.

Does it rewrite everything?

No, the model edits only flagged places.

Does it work in English?

Partly; the reference is Russian, and English detection is incomplete.

Related projects

jevsume
Resume review app and MCP server where a Cloudflare Worker asks Jev typed questions and composes a JevScore, generally or per job.
AppbetaTypeScript
zh-ste100-jev-skill
Chinese controlled-writing agent skill for technical reports with nine rules, using static checks plus a Jev semantic gate for certainty.
Agent skillbeta
jev-agent-skills
Portable skill pack giving agents a conservative Jev second opinion on intent, completion, routing and bounded browser navigation.
ProjectbetaPython
jev-flash-review
Local MCP review engine where each rule becomes a Jev YES/NO/N/A Choice over agent-supplied diff chunks, with located evidence.
MCP serverbetaTypeScript
jev-playwright-mcp
Drop-in proxy for Playwright MCP that adds Jev page triage, prompt-injection masking, snapshot pruning and risky-action gating.
MCP serverbetaTypeScript
jev-security-scan
Reviews agent skills, MCP configs and source with local static checks plus Jev before install, reporting evidence, risk and coverage gaps.
MCP serverbetaPython