DraftCheck AI-writing linter

io.github.kburrus64-maxv0.1.1更新于 Oct 5, 2026

DraftCheck: flags AI-writing patterns in prose with fix hints. Deterministic linter, not a detector.

已验证Streamable HTTP可网页运行Developer ToolsAI & ML

概览

AI 生成的概览

为文本中的常见 AI 写作模式打分，返回 0-100 分并给出逐条修改提示。

功能
DraftCheck 会扫描文本中 20 多种机器写作习惯，包括“不是 X，而是 Y”式对比、戏剧化的一句话结尾、夸大重要性、AI 常用词、破折号滥用、加粗标签式项目符号列表，以及“希望这对你有帮助！”之类的聊天机器人残留。它返回 0-100 分、类似“读起来像人写的”或“纯 AI 味”的标签，并指出每一处匹配及简短修改提示。它是确定性的检查工具，不是作者身份检测器，同一文本始终得到相同分数。
适用场景
适合在撰写或审阅文档、博客文章、发布说明等散文时，在发布前快速、可重复地检查读起来像 AI 生成的措辞。也适合编辑和智能体在写作步骤之后做一次确定性检查。
运行要求
使用提供方托管站点上的远程 MCP 端点；未声明身份验证、环境变量或请求头。另有网页应用和免费 JSON API，长文档和批量处理为按次付费端点。单独的 CLI 和库需要 Node.js 18+。
安装前请注意
托管 API 会把你的文本发送给第三方；长文档或批量的按次付费端点会产生费用。该工具标记的是写作模式而非作者身份：人类写作也可能得分很低，分数干净也不能证明是人写的。目前仅支持英文。

安装

在 SourceWeft 中

  1. 打开 控制台中的 DraftCheck AI-writing linter，将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Web executable，通过 Streamable HTTP。 远程服务在工作区中配置后即可从网页运行时运行。

其他 MCP 客户端

把它添加到你客户端的 mcpServers 配置中。

{
  "mcpServers": {
    "slopscore": {
      "type": "http",
      "url": "https://slopscore-nine.vercel.app/mcp"
    }
  }
}

README

DraftCheck

Formerly SlopScore. Same product, new name after a naming clash.

[test] [License: MIT] [Live site]

Find the patterns that make writing read as AI-generated, and fix them before you publish.

DraftCheck scans text for 20+ habits of machine-written prose: "it's not X, it's Y" contrasts, dramatic one-line closers, "let's dive in" openers, inflated significance ("a pivotal moment"), AI vocabulary ("delve", "seamless", "leverage"), em-dash overuse, bold-label bullet lists, and chatbot leftovers like "I hope this helps!". It returns a 0-100 score and points at every match with a short fix.

It's a linter, not a detector. It doesn't guess who wrote a text. It shows you the specific sentences readers will notice.

CLI

sh
npx github:kburrus64-max/draftcheck README.md docs/

(Not on npm yet as draftcheck. The unscoped slopscore name on npm belongs to a different project. Install from GitHub: npm i -D github:kburrus64-max/draftcheck. The slopscore CLI bin remains as an alias.)

FAIL  71  Pure slop        docs/launch-post.md           1: Chatbot residue: "Great question"           3: Not X, but Y: "this isn't just a tool, it's"           3: Inflated significance: "marks a pivotal moment"ok     4  Reads human      docs/install.md

Options:

flagdefaultwhat it does
--max <n>40exit 1 if any file scores above n
--format text|json|githubtextgithub prints annotations for Actions
--ext <list>.md,.mdx,.txt,.html,.rstextensions to scan inside directories
--ignore <ids>noneskip patterns, e.g. dash,triad
--include-quotedoffalso flag text inside double quotes (quoted examples are skipped by default)
--min-words <n>20skip very short files

Code blocks, inline code, front matter, URLs and HTML tags are ignored, so READMEs with examples don't get flagged for their code. Read from stdin with -.

GitHub Action

yaml
name: DraftCheckon:  pull_request:    paths: ["**/*.md", "docs/**", "content/**"]jobs:  slop:    runs-on: ubuntu-latest    steps:      - uses: actions/checkout@v4      - uses: kburrus64-max/draftcheck@v1        with:          paths: "docs content README.md"          max: "40"

Each match shows up as an annotation on the changed line. The check fails when a file goes over max.

Agent skill

DraftCheck ships as an Agent Skill for Claude Code, Cursor, Codex and other agents that read SKILL.md files. The skill tells the agent to score its draft, rewrite only the flagged spans, and re-check:

sh
npx skills add kburrus64-max/draftcheck

As a Claude Code plugin (skill plus the hosted MCP server), from inside Claude Code:

/plugin marketplace add kburrus64-max/draftcheck/plugin install draftcheck@draftcheck

Cursor can load the same repo as a plugin through .cursor-plugin/plugin.json.

Prompt-only writing skills such as humanizer and no-ai-slop tell an agent what to avoid; DraftCheck gives it a deterministic check to run afterwards.

Library

js
import { analyze, PATTERNS } from "draftcheck";
const r = analyze("Great question! Let's dive in.", { ignoreQuoted: false });r.score;    // 0-100r.label;    // "Reads human" | "A little sloppy" | "Sloppy" | "Pure slop"r.matches;  // [{ id, label, category, start, end, text, hint }]

No dependencies. Works in Node 18+ and the browser.

How scoring works

Each match has a strength. Strong tells (chatbot leftovers, not-X-but-Y, dramatic closers) count 3 points, medium tells 2, weak tells 1. Points are divided by text length (per 100 words, with a 100-word floor) and mapped to 0-100. Weak signals like a single em dash or one three-item list don't count unless stronger tells are also present, because careful human writers use them all the time.

The rules are regular expressions plus a sentence-shape check for rows of fragments. They are deterministic: the same text always gets the same score.

Limits

  • English only for now.
  • It flags patterns, not authorship. Plenty of human writing has a few of these, and a clean score doesn't prove a person wrote something.
  • The CLI skips text inside double quotes, so style guides can quote bad examples. The web app and API count quoted text unless you pass ignoreQuoted.

Hosted API

The web app has a free JSON API (POST /api/check, up to 5,000 characters), a URL scorer, and pay-per-call endpoints for longer documents and batches. See https://slopscore-nine.vercel.app/llms.txt.

Credits

The pattern list builds on Wikipedia's Signs of AI writing (WikiProject AI Cleanup) and two MIT-licensed agent skills: blader/humanizer and petergyang/no-ai-slop. The detection code and scoring here are original.

Notes on the rename

This project was formerly SlopScore. The GitHub repo is now kburrus64-max/draftcheck (old URL redirects). The MCP Registry namespace stays io.github.kburrus64-max/slopscore so the existing listing is not orphaned. Live site and MCP remote remain at https://slopscore-nine.vercel.app.

License

MIT © Anansi Data. See LICENSE.

来源：README.md，提交 5cefa88

工具

0
工具元数据尚未被收录。

版本历史

1
  1. v0.1.1最新Oct 5, 2026