repo·evals
· 2026-04-24

Humanizer-zh

op7418/Humanizer-zh

🛠63 / 100
✅
⚠
🎯
⚠

📝
🗺
01Signal scanning信号发现02Content acquisition内容获取03Content understanding内容理解04Topic curation选题决策05Content production内容生产06Creative assembly创意组装07Distribution & feedback分发反馈08Learning学习
📍
⚛
→
⚗
→
🧬

🛑
0–29
⚠️
30–49
🛠
50–79
🏭
80–100
▼
63
🛠· 63 / 100
  • ✓7 claims passed, no critical failures
  • ✓MIT / Apache / etc., installable per deployment.install_methods
  • ⚪no recent release pipeline + not recently active
  • ⚪EN-only or ZH-only README
  • ◐evidence_completeness=portable

#1👤
#2🎯
#3🧭
#4⇄

Chinese AI-suspect draft中文 AI 嫌疑草稿(800-2000 chars)(800-2000 字)Humanizer-zh skillHumanizer-zh skill(detect → rewrite → score)(识别 → 改写 → 打分)Humanized text去 AI 味文本+ 50-point score+ 50 分打分

npx skills add op7418/Humanizer-zhany (npm)easy
git clone into ~/.claude/skills/humanizer-zhanyeasy
  • 📡
Anthropic Claude / Claude Code
LLM that runs the humanization rubric
Standard agent-side cost; pure prompt-skill, no external services
· 7
7
+40
+23
0
0
0
0

7 / 7
passed claim-001

passed claim-002

passed claim-003

passed claim-004

passed claim-005

passed claim-006

input_contract
output_contract
determinism
idempotence
no_skill_callouts
failure_mode_clarity

  • evidence_completeness='portable' → capped at 'reusable'

archetype: prompt-skill→core_layer_tested? True→evidence: portable→recommended: reusable→final: reusable
ceiling 1 · evidence_completeness='portable' → capped at 'reusable'
claim-001SKILL.md frontmatter 合法且完整criticalskill-structure● passed
claim-002无悬空内部引用criticalskill-integrity● passed
claim-003README 承诺的 24 种 AI 模式 SKILL.md 全部覆盖criticalcoverage● passed
claim-004真实会话：能识别并改掉 AI 痕迹criticalend-to-end-llm● passed
claim-005自评分系统能用highend-to-end-llm● passed
claim-006安装说明路径至少有一条可用highinstall● passed
claim-007不是简单的英译中——有中文适配mediumlocalization● passed

0%
0.00s
0

# Final Verdict — op7418/Humanizer-zh

## Repo

- **Name**: op7418/Humanizer-zh
- **Version tested**: commit 91f3d39 (cloned 2026-04-24)
- **Date**: 2026-04-24
- **Archetype**: prompt-skill
- **Final bucket**: 🟡 reusable
- **Confidence**: medium

## Plain English

- **Outcome if adopted**: 把中文 AI 稿扔给它，能识别 24 种 AI 写作痕迹，真的改成更像人说的话；自带 50 分评分能量化判断改得好不好。
- **Regret scenario**: 它改的不只是文风——内容会变短，AI 编出来"看起来对但其实没有依据"的部分会被直接删掉。发稿前没复核事实就可能漏掉。

## Why This Bucket

- 封顶理由：`evidence_completeness='portable'` → 最高到 `reusable`
- 未到 `recommendable`：只跑了 1 次 live run；README 首选的 `npx skills add` 安装路径依赖未验证的外部 CLI
- 未到 `usable` 以下：核心 live run 通过（13 处痕迹全识别、正文中 17 个黑名单词汇 0 命中）、静态结构干净、中文本地化真实

## Evidence

| claim | priority | status | 说明 |
|-------|----------|--------|------|
| claim-001 frontmatter 合法 | critical | passed | name/description/allowed-tools/metadata.trigger 齐全 |
| claim-002 无悬空引用 | critical | passed | 仅 SKILL.md + README.md + LICENSE，自包含 |
| claim-003 24 模式全覆盖 | critical | passed | README 4×6 模式在 SKILL.md ### 1–24 全部对应 |
| claim-004 live run 去痕成功 | critical | passed | 13 处 AI 痕迹全识别，改写正文黑名单命中 0 |
| claim-005 50 分评分可用 | high | passed | 本次 44/50，5 维每维给出具体理由 |
| claim-006 至少一条安装路径可用 | high | passed_with_concerns | git clone 路径必然可用；npx 路径未验证 |
| claim-007 中文本地化 | medium | passed | 「」引号注释、中文 AI 词汇、中文示例 |

## What Would Take This to 🟢 Recommendable

1. 再跑 ≥ 2 次独立 live run，覆盖不同文体（营销/学术/博客）
2. 验证 `npx skills add` 这条 README 首选的安装路径到底能不能跑
3. 给评分档位做个一致性测试（同一段文本，两人 / 两次分别打分差距 ≤ 5 分）

## Artifacts

- `areas/end-to-end-llm/runs/2026-04-24/run-live-humanize-1/artifacts/01-input.md`
- `areas/end-to-end-llm/runs/2026-04-24/run-live-humanize-1/artifacts/02-output.md`
- `areas/end-to-end-llm/runs/2026-04-24/run-live-humanize-1/logs/anti-slop-grep.log`
- `verdicts/2026-04-24-verdict.html`（用户看到的产品页）