Mc Ai Tell Audit logo

Mc Ai Tell Audit

Community
jtydhr88
mc-ai-tell-audit

The post-generation AI-tell audit (AI 味诊断). Run this on every generated track before accepting it. Covers the fourteen enumerable defaults that generative music models fall into, which of them are machine-measurable and which need ears, why compliance rate and AI-tell count are two separate numbers that must never be merged, the zero-point calibration that sets the pass threshold at 82 rather than 60, the ordered diagnostic path from symptom to owning skill, and the remediation list. Use when a generation comes back and must be accepted or rejected, when a track sounds synthetic but you cannot say why, when deciding whether to accept a take, or when a high compliance rate still produced a bad result. AI 味、诊断、验收、听不出来是 AI、照做率、生成后检查。

Overview

Publisherjtydhr88
Repositorymusic-composition-skills
Skill namemc-ai-tell-audit
Stars
171
Forks
23
Bundled files
Instructions only
LicenseMIT
Links
  • Markdown instructions

    A SKILL.md file the model loads on demand, so it only costs tokens when a request actually matches.

  • Works with any LLM

    AI skills are plain Markdown, not provider-specific code, so this works with GPT, Claude, Gemini, Grok, or a local model.

  • Self-contained

    Everything the model needs lives in the instructions — no extra files to sync.

  • Open source

    Published by jtydhr88 on GitHub. Read the source before you install it.

Installation

Install the Mc Ai Tell Audit AI skill in TypingMind to use it with any LLM, or drop it into another agent that reads SKILL.md.

1

Install in TypingMind

TypingMind installs a skill straight from its GitHub folder — it reads SKILL.md, bundles the resource files, and stores the result locally.

  1. Open the app and go to Plugins → Skills.
  2. Choose "Install from GitHub".
  3. Paste the skill folder URL below and confirm.
    https://github.com/jtydhr88/music-composition-skills/tree/main/plugins/music-composition/skills/mc-ai-tell-audit
  4. Enable the skill in any chat where you want it available.
Plugins → Skills → Add skill → From GitHub URL, then paste the folder URL and press Continue.
2

Install in another agent

Any agent that reads the Agent Skills format can use this skill — copy the folder into that agent's skills directory.

Claude Code — .claude/skills
git clone --depth 1 https://github.com/jtydhr88/music-composition-skills.git /tmp/music-composition-skills
mkdir -p .claude/skills
cp -r /tmp/music-composition-skills/plugins/music-composition/skills/mc-ai-tell-audit .claude/skills/mc-ai-tell-audit
Restart Claude Code after copying so it picks up the new skill.

Use it in TypingMind

Enable Mc Ai Tell Audit in any TypingMind chat and the model takes it from there. Its name and description sit in the system prompt, and the moment a request matches, the model loads the full instructions itself — you never invoke it by hand, and it costs no tokens until it is actually used.

The model loads Mc Ai Tell Audit on its own as soon as a request matches it.

Works with any AI model

AI skills are plain Markdown instructions rather than provider-specific code, so Mc Ai Tell Audit is not tied to the model it was written for. Install it once in TypingMind and use it with GPT-5, Claude, Gemini, Grok, DeepSeek, Mistral, Llama, or a local model you run yourself — all on your own API keys.

  • Loaded only when it is needed

    The system prompt carries just the name and description. The instructions are fetched on the first matching request, so an idle skill costs nothing.

  • Switch models mid-chat

    Because the skill is instructions rather than code, changing model does not break it — the next model reads the same SKILL.md.

Skill instructions

This is the SKILL.md content the model loads. Read it before installing — a skill is instructions your model will follow.

AI 味诊断(AI-Tell Audit)

生成回来的每一条都要过这个 skill。这是本库的终检。

核心命题:AI 味不是玄学。 生成模型不给规格就走默认值,而这些默认值是可枚举的。 本 skill 就是那份枚举,加上"每一条怎么查、归谁管"。

按任务读哪几节

任务读
刚生成完,要验收§2 十四条 → §3 判定
"听着像 AI 但说不出哪里"§2(★ 逐条查,不要凭直觉)
照做率很高但结果不好§4
要定阈值§3.2 零点校准
查出问题了要改§5 整改路径
人声的四条怎么查§6

边界

不归这里归哪
规格写完时的检查mc-workflow §3.0 的 20 条自查表
照做率怎么算本 skill §3.1(定义与标定);honors 三档怎么影响计分见 mc-render-compile §6
具体怎么改各 L1/L2 skill(§5 给路由)
后端做不到的字段怎么办mc-render-compile §6、§7

★ 与 20 条自查的分工: 自查问"规格写完没有",本 skill 问"生成出来的东西像不像人做的"。 两边都看过才算完。 17 条全过的规格照样可能生成出满是 AI 味的音频。


1. 动笔前必填

必填问法
① 这一版的规格与种子没记录就没法复现,审计结果也就没用
② 后端的 honors 表哪些项是它做不到的——那些不算它的账
③ 有没有分轨人声四条(§6)需要先分离人声轨

2. ★ 十四条枚举

逐条打勾。不要凭整体印象。

#AI 味来源(后端默认值)对抗的 spec 字段可自动测归谁管
1段落全是 8 的整数倍form[].bars✅ 段落边界检测mc-arrangement-arch §4
2能量曲线单调递增,从不减energy_curve、subtraction_events✅ 分段 RMSmc-arrangement-arch §2、§3
3编制从头到尾不变roster[].entry/exit✅ 分轨 onset 密度mc-arrangement-arch §5
4全曲一个和声循环到底bridge 必须离调或转调✅ 和弦识别mc-harmony、mc-modulation
5fill 刻板落在 4 的倍数fill_policy✅ onset 峰值位置mc-rhythm-section §8
6没有非人声记忆点hooks.arrangement_hook⚠️ 半自动(重复段落相似度)mc-arrangement-arch §7
7动态压死、频谱质心恒定dynamics、width_map✅ DR/LRA、质心时序方差mc-mix-intent
8网格化、无 push/pullgroove.push_pull✅ onset 相对网格偏移mc-rhythm-section §3
9时长永远 3:30,没有尾巴target_duration + 结尾处理✅mc-arrangement-arch §8
10每句同样的力度与咬字,没有气声、破音、rubatovocal.delivery、dynamics_by_section⚠️ 需分轨后测短时响度方差mc-vocal-direction §2、§4
11副歌三遍唱得一模一样vocal.chorus_variation✅ 人声轨分段相似度mc-vocal-direction §1④
12换气点不自然或听不到换气vocal.breath_points⚠️ 半自动mc-vocal-direction
13ad-lib 均匀撒或干脆没有vocal.ad_libs⚠️ 半自动mc-vocal-direction
14音准全中vocal.pitch_policy⚠️ 半自动(分轨后测音高偏移分布)mc-vocal-direction §4

★ 这张表一张三用:本 skill 的骨架、AB 实验评判表的骨架、 以及 20 条自查的来源。改一处要同步三处。

2.1 ★ 三条最容易漏、也最致命的

不是因为它们最难查,而是因为它们不难听——只是"不像人做的"。

#为什么致命
2(能量只升不降)实测 62/62 真实编曲都有下降。这是命中率最高的一条强规则,也是听感上最像 AI 的一条
8(全部对齐网格)实测 62/62。注意读法是"不许全曲每一件都在网格上"——鼓组绝对量化是允许的(→ mc-rhythm-section §2.0)
6(没有非人声记忆点)它的症状是"听着还行但记不住",最容易通过所有其他检查。缺它的歌听完就忘

3. 两个数,不许合并

3.1 定义

问什么怎么得范围
照做率后端照着规格做了多少逐字段对照规格0–100
AI 味旗标数成品有多少条机器指纹本 skill §2 逐条打勾0–14

★ 绝对不要把它们合成一个分数。 它们诊断的是不同的病:

照做率高 + 旗标少  → 好,接受
照做率高 + 旗标多  → ★ 规格本身写得像 AI(见 §4)
照做率低 + 旗标少  → 后端自己干得不错,但你不可复现(见 §4.2)
照做率低 + 旗标多  → 编译或后端选择有问题,回 mc-render-compile

3.2 ★ 零点校准:阈值为什么是 82 不是 60

不相关的、制作精良的成品在照做率表上得 63–66 分。

这是因为任何一首正常的歌都会"碰巧"满足一部分项(有拍速、有段落、有动态)。 所以:

线值
零点(不相关成品)63–66
及格70
合格82

★ 上面三个数是参考值:零点是拿一份 City Pop 规格(112 BPM、10 段、114 小节)去比三首与它无关、制作精良的华语成品,按字段加权打分测的(63.4–66.4;同一规格比一段纯 MIDI 渲染的 demo 只有 39.5)。用来理解量级,不是交付门槛;逐字段对照时,照做率就是对照后的一个判断。

★★ 一个踩过的坑:把权重从有区分度的项(boundaries)挪到 没区分度的项(energy_shape,不相关曲目也能拿 0.8), 会抬高所有人的分数,包括零点基线(66→72)。 加权重要加在能区分的项上。

3.3 后端做不到的不算它的账

honors: none 的字段排除计分,不是给 0 分。 详见 mc-render-compile §6.1——这是一个正确性问题,不是口径问题。


4. ★ 照做率高但仍然像 AI

这是本 skill 存在的主要理由。

照做率只衡量"后端有没有照做",不衡量"规格本身好不好"。 规格写得像 AI,后端忠实执行,结果就是一首照做率 95 的 AI 歌。

4.1 规格像 AI 的四个征兆

征兆查哪
intent.one_thing 是为了过 lint 填的,念出来不像一句人话mc-workflow §1 的 S0
reference_pair 的 borrow 写得很泛("借它的氛围")同上。"只借前奏的空"才叫写了
每个段落的 bars 只是为了避开 8 的倍数而选的怪数字mc-arrangement-arch §4——非对称要有理由
用的进行 / 转调命中了套路警戒线且没写理由mc-progressions §3、mc-modulation §6

★ 共同点:规则被当成了要满足的条件,而不是要做的选择。 这正好是 AI 味的定义——没有人做过选择。

4.2 照做率低但好听

说明后端自己发挥得不错。可以接受这个成品,但要意识到:

  • 你不可复现它
  • 你不知道是哪个决定让它好
  • 下一首得从头碰运气

★ 处理办法:把这个成品逐字段反推成一份实测 ARR-SPEC(边听边填,测不准的标"估算"), 看它和你原本的规格差在哪。那个差值就是你学到的东西。


5. 整改路径

按这个顺序,不要凭直觉跳。

听着"像 AI"
 └─ ① 逐字段对照规格,拿照做率
 └─ ② 逐条过 §2 的十四条,数出旗标数
     ├─ 照做率低  → 查 honors
     │   ├─ 这个字段该后端是 none → 不是它的错。换后端(YuE2 控制力最强)或接受
     │   └─ 不是 none → 编译问题 → mc-render-compile §3
     └─ 照做率高但旗标多 → ★ 规格本身的问题 → §4.1

5.1 症状 → skill(速查)

症状先查
平、没起伏mc-arrangement-arch §2
糊、浑mc-texture-layering §4(★ 先查持续音,不是先怪鼓)
主角不突出mc-texture-layering §3 四条深度线索
机械、太准mc-rhythm-section §2 + mc-vocal-direction §4
听着还行但记不住mc-arrangement-arch §7(arrangement_hook 是不是空的)
第二遍副歌和第一遍一样mc-arrangement-arch §6 + mc-vocal-direction §1④
人声埋在伴奏里先查音准(mc-vocal-direction §4.1),再查频段
贝斯在手机上消失mc-sound-design §6 差音
结尾突兀 / 淡出了事mc-arrangement-arch §8;★ 检查 exclude 里有没有排除 fade out ending
技术上干净但情绪不对mc-mix-intent §3.1(mood 是不是没写)
和声一个循环到底mc-harmony §7、mc-modulation

6. 人声四条(#10、#12、#13、#14)

这四条都需要先把人声轨分离出来。

6.1 分轨

用任何分轨工具把人声轨分出来即可(demucs 一类,或用户生成载体自带的分轨,通常分成人声/贝斯/鼓/其余)。 本 skill 不预设哪一个。

★ 这是分轨在本库里少数必要的场合之一: 多轨语料(Cambridge-MT)给不了"Suno 唱出来的人声",只能从生成结果里分。

6.2 分轨之后测什么

#测什么判据
10 力度人声轨的短时响度方差方差接近 0 = 全程一个力度
11 副歌雷同各遍副歌人声轨的分段相似度相似度过高 = 原样复制
12 换气句间的低能量段⚠️ 半自动,目前靠听
13 ad-lib装饰性发声的分布⚠️ 半自动。均匀 = 撒的,不是唱的
14 音准音高偏移的分布⚠️ 半自动。全部落在 ±5 音分内 = 修过头/生成的

6.3 ⚠ 现状

#12、#13、#14 靠人工听,这是刻意的选择,不是漏了一步。

这三条测的是换气的能量谷、装饰音的分布、音高偏移的分布,它们的自动判据都要按生成模型分别标定阈值,不同模型的基线差得远,硬套一个统一数字容易把唱得好的判成有问题。人工听一遍这三条本来就很快,也正是耳朵天然比机器灵的地方,不算额外负担。


7. 验收清单

每一条生成结果都要过。

  • 记录了规格版本与种子
  • 逐字段对照过规格,照做率心里有数
  • 逐条过了 §2 的十四条,数出了旗标数(不是凭印象)
  • honors: none 的项被排除计分,不是给 0
  • 照做率高但旗标多的话,查过 §4.1 的四个征兆
  • 人声四条至少人工听过一遍
  • ★ 耳朵是最后介入的:A/B 打乱 → 成对比较 → 同一对听两遍看结论稳不稳
  • 结论两遍不一致 → 这个差异不存在,别改

附:来源

  • 实测数据(§2.1 的 62/62、下面的 22% 与 5/7):在 Cambridge-MT 多轨 62 首自然抽样上统计;5/7 是其中 7 首人工标注段落边界的真值
  • 零点 63–66:见 §3.2
  • 本 skill 的十四条全部可以人工判
  • 十四条里 1–5、7–9、11 有音频分析就能机器算,6、10、12、13、14 是人工项

★ 一条方法论提醒(踩过的坑): 指标测不准时,不要拿指标去推翻教科书,先问指标测的是不是那回事。 例:「最高潮前应有能量回落」在全量 62 首里只有 22%,但在 7 首人工标注真值里是 5/7—— 差距来自算法用能量峰值定位"最高潮",而人耳不是。 22% 那个数测的不是这条规则。

Frequently asked questions

What does the Mc Ai Tell Audit AI skill do?

The post-generation AI-tell audit (AI 味诊断). Run this on every generated track before accepting it. Covers the fourteen enumerable defaults that generative music models fall into, which of them are machine-measurable and which need ears, why compliance rate and AI-tell count are two separate numbers that must never be merged, the zero-point calibration that sets the pass threshold at 82 rather than 60, the ordered diagnostic path from symptom to owning skill, and the remediation list. Use when a generation comes back and must be accepted or rejected, when a track sounds synthetic but you can...

Why use Mc Ai Tell Audit on TypingMind?

Because you install it once and use it with any model. Mc Ai Tell Audit is plain Markdown rather than provider-specific code, so the same skill runs on GPT-5, Claude, Gemini, Grok, or a local model — and you can switch model mid-chat without it breaking. TypingMind runs on your own API keys, so you pay providers directly instead of a per-seat subscription, and your skills and chats stay in your own storage.

How do I install Mc Ai Tell Audit in TypingMind?

Open Plugins → Skills → Install from GitHub in TypingMind and paste https://github.com/jtydhr88/music-composition-skills/tree/main/plugins/music-composition/skills/mc-ai-tell-audit. TypingMind reads its SKILL.md and installs it as a skill you can enable per chat.

Which AI models can use Mc Ai Tell Audit?

Any model you connect in TypingMind. AI skills are plain Markdown instructions rather than provider-specific code, so GPT, Claude, Gemini, Grok, and local models can all load this skill when a request matches it.

How many AI models can I use with Mc Ai Tell Audit?

As many as you like. As long as a model supports skills, you can use Mc Ai Tell Audit with it — GPT, Claude, Gemini, Grok, DeepSeek, Mistral, Llama and more — all on TypingMind with your own API keys.

Is the Mc Ai Tell Audit AI skill free?

Yes. It is published on GitHub by jtydhr88 under the MIT license. You only pay your own AI provider for the tokens you use.

What are AI skills?

An AI skill is a reusable instruction bundle that teaches an AI model how to do one specific task. It follows the open Agent Skills format: a SKILL.md file with a name and description, plus any scripts, templates or reference files the model may need. The model reads the instructions only when your request matches the skill, so an installed skill costs nothing until it is used.

How are AI skills different from plugins or MCP servers?

A plugin or MCP server gives a model new tools to call — code that runs somewhere and returns a result. An AI skill gives the model knowledge and process instead: how to approach a task, which steps to follow, what good output looks like. Skills are plain Markdown, so they need no server, no API key and no runtime, and they work with any model.

View all

Set up your own AI workspace now

Get notified about new features and future giveaways by subscribing to our newsletter 👇