agentleFS
liush2yuxjtuliush2yuxjtu/douyin-video-skill/skills/verify-skill/SKILL.md

verify-skill

Skill0 starsChanged 32 days agoMITView on GitHub

Use before committing or pushing changes to this repo (douyin-video-skill) to run the local verification suite (there is no CI) — required files, SKILL.md frontmatter, .claude-plugin JSON, attribution of quoted creator content, broken relative markdown links, and bash -n / shellcheck on scripts. 提交或推送前运行本地校验(本仓库不使用 CI)。

1 other project uses a version of this.

Adopted by

Most recently taken up by

  • Other versions1 project uses a version of this.

    Word for word · 0

    None found.

    Edited · 0

    None found.

    Wrote their own version · 1

    1. NeuroAIHub/awesome_cognitive_and_neuroscience_skills★ 39same name

    How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.

  • In this repositoryOne other file it gives its agents.

The file

We split this skill into the rules it gives an agent

5 rules. About 0 are common, 2 are distinctive, and 3 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 3

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Every skills/*/SKILL.md has frontmatter with name (= directory, lowercase-hyphen) and a non-empty description (≤1024 chars), plus a body.

    line 19
  2. Distinctive:

    Every relative markdown link resolves to an existing file.

    line 22
  3. Very specific:

    When adding a skill from a new video, also add its source to ATTRIBUTION.md.

    line 25
  4. Very specific:

    Use 20–30 s of natural, continuous speech from the real speaker, same phone/room, normal pace.

    line 85
  5. Distinctive:

    Keep the original script text.

    line 89

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Local verification / 本地校验 (no CI)

This repo intentionally has no GitHub Actions. Run this locally before every push. 本仓库不使用 GitHub Actions,每次推送前请在本地运行。

Run / 运行

bash skills/verify-skill/verify.sh

Needs bash, python3; shellcheck is used if installed, otherwise skipped. Exit code 0 = all passed. 需要 bash、python3;装了 shellcheck 会自动使用。退出码 0 表示全部通过。

What it checks / 检查项

  1. Required files: README.md, README_zh.md, LICENSE, .claude-plugin/*.json; no .github/workflows.
  2. Every skills/*/SKILL.md has frontmatter with name (= directory, lowercase-hyphen) and a non-empty description (≤1024 chars), plus a body.
  3. .claude-plugin/marketplace.json and plugin.json are valid JSON, names and versions match.
  4. Attribution: ATTRIBUTION.md exists, mentions the source title, creator and link, and is linked from both READMEs and from every skill that ships a references/ transcript.
  5. Every relative markdown link resolves to an existing file.
  6. bash -n and shellcheck -S warning on every *.sh.

When adding a skill from a new video, also add its source to ATTRIBUTION.md. 新增来自其他视频的 skill 时,请同步在 ATTRIBUTION.md 中补充出处。

A-roll 提词录制器端到端验证 / Teleprompter E2E verify

Repo: https://github.com/liush2yuxjtu/aroll-teleprompter · Site: https://aroll-teleprompter.vercel.app/?k=<TP_KEY>&script=<id>(koubo / chat20 / voice-sample)

git clone https://github.com/liush2yuxjtu/aroll-teleprompter && cd aroll-teleprompter
npm ci && npx playwright install chromium
TP_KEY=<key> node verify/verify.mjs        # 可选 BASE=<预览地址>;默认生产地址

Headless Chromium with --use-fake-device-for-media-stream, mobile (Pixel 7) + desktop (1440×900) viewports, for every script in public/scripts/index.json: 标签数量与切换、顶部新手说明、逐句字幕(#cur 与 n / 总数)、录制约 7s → 等「已自动上传」→ GET /api/recordings?session= 为 complete、script 字段正确、分片 ≥1、markers.json 的 script.id 正确; 最后只 DELETE 本次测试创建的会话(绝不删除真实录音)。通过标准:输出 RESULT: ALL PASS、退出码 0。改提词站后必须对生产地址跑通再报告完成。

Deploy note: the site is not Git-connected — deploy via Vercel MCP upload_file + create_deployment (target: production). Vercel MCP: omit teamId (403 bug) — passing teamId/slug returns 403 forbidden; omit them. Voice comparison page: /voices/?k=<TP_KEY>, read the pick via GET /api/choice?k=<TP_KEY> (.latest.choice). 部署说明:站点未连 Git,用 Vercel MCP 上传文件后 create_deployment;不要传 teamId/slug(会 403)。

手机文字显示检查 / Mobile text-display checks

verify/verify.mjs 先跑这部分(TEXT_ONLY=1 只跑它,不上传):iPhone SE 375×667、iPhone 14 390×844、Pixel 7 412×915、小屏安卓 360×740 × 竖屏/横屏 × 空闲/录制中布局 × 每个脚本的每一句:

  • 无横向溢出:scrollWidth <= clientWidth
  • 字幕 #cur 完整在视口内,且文字区域 elementFromPoint 命中的是字幕自身(没被按钮/控制栏遮挡)
  • 字号在 16–72px;字幕框不裁切:scrollHeight <= clientHeight
  • 中文非豆腐块:canvas measureText 宽度与缺字字形对比(box 上需装 Noto Sans CJK,避免误报)
  • 截图存 verify/artifacts/(ART= 可改)。用户反馈“手机上字显示不对”时,先跑这部分确认能复现(应 FAIL),修完再跑到 ALL PASS。

克隆配音校验 / Voiceover (voice-clone) verify

Script: verify_voiceover.py. Needs ffmpeg, numpy, faster-whisper (CPU, compute_type=float32; audio decoded via ffmpeg → float32 numpy, no torchaudio). 需要 ffmpeg、numpy、faster-whisper(CPU、float32;用 ffmpeg 解码成 numpy 送入)。

python3 skills/verify-skill/verify_voiceover.py \
  --full vo_full.wav --clips final/ --lines lines.txt --ref ref_real.wav \
  --nouns 杜威,默顿,波利亚,置换        # 专有名词常被 ASR 识错,相似度计算时忽略

--clips = one wav per sentence, sorted by filename in line order; --lines = script, one sentence per line. --clips 每句一个 wav(按文件名排序 = 台词顺序);--lines 每行一句。

Checks / 检查项 (PASS / WARN / FAIL table, exit 1 on any FAIL):

  1. Text — per sentence + full file, char similarity (digits → 汉字, punctuation stripped, --nouns removed): <0.80 FAIL, <0.90 WARN. Proper nouns and homophones (长→找, 教程→叫常) are often ASR errors, so listen to WARNs before regenerating; a repeated mis-hear of a common word (e.g. 先→卸) is more likely a real slip. 逐句 + 全文字符相似度;WARN 先人工听,常见词被反复识错更可能是真读错。
  2. Pitch — median f0 of the voiceover vs the real-voice reference within ±10% (catches the "announcer" pitch-up: old clone 164 Hz vs real 116 Hz). 音高中位数与真人参考相差 ≤10%(播音腔会整体偏高)。
  3. Rate — chars/s over speech within the reference's per-segment range ±20%. 语速(字/秒)在真人参考范围 ±20% 内。
  4. Gaps / junk — max silence in full file ≤1.5s; per clip: duration ≤ chars/2 + 2.5s and inner silence ≤1.0s. Catches the TTS reading the style prompt aloud (seen on very short lines like 「下期见。」) and long dead air. 全文最长静音 ≤1.5s;单句时长与内部静音上限——可抓到 TTS 把风格提示词念出来、长空白。
  5. Level — true peak ≤ −1 dBFS, integrated −14 ± 1.5 LUFS; a raw clip at 0 dBFS is a WARN (source take may be clipped, prefer regenerating). 真峰值 ≤ −1 dBFS,响度 −14±1.5 LUFS;单句原始峰值 0 dBFS 给 WARN。
  6. Prosody (phrasing + intonation) — needs jieba (else WARN-skip; --no-prosody to turn off). Full-file word timestamps are aligned to the script; each energy pause (≥0.08s, −30 dB) is snapped (±1 char, ASR jitter) to a script boundary: sentence end / comma / jieba word boundary / inside a word. 韵律(断句 + 语调):需要 jieba。把每个停顿对齐到台词边界:句末 / 逗号 / 词边界 / 词内。
    • FAIL run-on: fewer than 30% of commas get a pause (--min-comma-pause) — sounds breathless/odd. 逗号处几乎不停顿 = 一口气念完,听感断句怪。
    • FAIL mid-phrase break: any pause ≥0.15s not at punctuation (--max-midphrase), e.g. 可能是|三个月前. 非标点处停顿 ≥0.15s。
    • FAIL inverted hierarchy: a non-sentence-end pause ≥0.9× the mean sentence-end pause (comma pauses must be clearly shorter than full stops). 逗号停顿不应接近句末停顿。
    • WARN flat intonation: mean per-sentence f0 range < 7 semitones (--min-f0-range); phrase-final moves are printed (clear falls ≈ assertive/"good tone"). 每句音高范围 <7 半音给 WARN。 Calibrated on a human listening test (3 style samples): run-on A and mid-phrase-break B were judged 「断句太奇怪」, C 「连贯」; B (13 st range, −3.5 st final falls) 「语气语调好」. The metrics reproduce all three. Root cause there was the TTS take itself (raw takes already had 0/9 comma pauses, or 1.2 s mid-sentence gaps and 可能是|三个月前); silence-shortening only acts on <−32 dB gaps, so it cannot splice voiced audio, but it flattens pause contrast — re-take bad sentences instead of over-trimming. 以 3 个风格样本的人工听感校准:A(一口气)和 B(词中断)被评「断句怪」,C「连贯」,B「语气语调好」,指标全部复现。根因是 TTS 原始 take,而非剪静音/变速;坏句应重生成。

Clone reference rule / 克隆参考音规则

  • Use 20–30 s of natural, continuous speech from the real speaker, same phone/room, normal pace. Record it with the teleprompter scripts chat20 (casual chat) and voice-sample (clean read); trim only leading/trailing silence, light normalization, no EQ/reverb. 用真人 20–30 秒自然连续说话(同一手机、同一房间、正常语速),用提词器脚本 chat20 / voice-sample 录制;只裁首尾静音、轻度归一化,不加 EQ/混响。
  • A test clip like 「喂喂喂…」 fails: it carries timbre but no tones/prosody, so the model falls back to a polished announcer delivery (uncanny, pitch too high). Do not compensate with fake reverb or dashes/ellipses in the text — they add dead air. 「喂喂喂」式参考音会失败:只有音色没有声调与语调,模型退回播音腔。不要用假混响或在文本里加破折号/省略号来补救。
  • Keep the original script text; style prompt casual (「像平时跟朋友聊天一样随口说……不要播音腔」); 2 takes per bad sentence, pick by this script. 文本保持原稿;风格提示要口语化;问题句各生成 2 条,用本脚本挑选。

verify.sh only byte-compiles this script (no audio needed); set VO_DIR=<dir with vo_full.wav, final/, ref_real.wav, lines.txt> to also run it (skipped if faster-whisper is missing). verify.sh 默认只做语法检查;设置 VO_DIR= 才会真正跑配音校验(缺依赖则 SKIP)。

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.