agentleFS
full-aigc-skillsfull-aigc-skills/minimax-skills/skills/music-video-subtitle-generator/SKILL.md

music-video-subtitle-generator

Skill4 starsChanged 59 days agoApache-2.0View on GitHub

For musicians, video creators, and social-media editors producing AI music videos or emotional short films with lyric typography. Users provide music, lyrics, references, characters, typography direction, mood, or target platform. The Skill analyzes beat and vocal timing, separates character, scene, and text references, designs beat-reactive spatial typography, decomposes long works into connected shots, audits prompts, and routes generation for H3 or other video tools. It outputs MV concepts, shot prompts, lyric text plans, and stitching guidance. Best for stylized MVs and subtitle-driven music visuals, not ordinary caption cleanup, licensed IP copying, or fully manual post-production editing.

  • Installs packages

9 other projects use a version of this, 5 of them edited.

Adopted by

Most recently taken up by

The file

Show its rules and text

We split this skill into the rules it gives an agent

23 rules. 20 are common, 0 are distinctive, and 3 are very specific to this project.

  • 20

    Common

    Most of the 7 edited versions we read keep these.

  • 0

    Distinctive

    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. Common:

    Do not use it for ordinary subtitle burn-in, generic video editing, non-music product ads, or simple single-image / single-clip requests …

    line 36
  2. Common:

    Do not assume third-party tools, shell scripts, browser plugins, or external generation APIs.

    line 40
  3. Very specific:

    Write locked prompts and revisions to the episode ledger (complete-mv-prompt.md, decisions.md).

    line 41
  4. Common:

    Do not hardcode output paths.

    line 43
  5. Common:

    If the user says “skip confirmation / just do it”, apply it only to the current run unless the outer orchestrator has active session …

    line 44
  6. Common:

    If the user only wants a prompt, stop after delivery.

    line 45
  7. Common:

    Use the latest confirmed creative intent.

    line 50
  8. Common:

    Do not mechanically add performance, typography, transitions, camera moves, or character actions just because a preset includes them.

    line 52
  9. Common:

    If the user uploads a real song or beat, it is the master music bed unless they request replacement.

    line 54
  10. Common:

    Before writing the final production prompt, confirm the minimum foundation with concise recommended choices.

    line 60
  11. Common:

    Before prompt writing or video generation, confirm target MV duration.

    line 78
  12. Common:

    If uploaded music exists, the confirmed duration is the music window length.

    line 87
  13. Common:

    If uploaded or specified music is longer than the confirmed duration, lock a segment before writing prompts:

    line 91
  14. Common:

    If music is shorter than the target duration, use the full audio unless the user explicitly asks to stretch or extend it.

    line 97
  15. Common:

    Before final video prompts or generation, determine lyric ownership:

    line 101
  16. Common:

    If the user provides lyrics, those are locked lyrics.

    line 103
  17. Common:

    If the user wants a finished music-aesthetic MV without lyrics, generate short original lyrics matching the music style, vocal mode, …

    line 104
  18. Very specific:

    Every Typography: field must come from the same locked lyrics.

    line 107
  19. Common:

    Before final prompts, create a concise creative contract:

    line 121
  20. Common:

    Keep scene mood unified, but each shot’s space must differ clearly, such as local background, tunnel, wall, light strip, or reflective …

    line 142
  21. Common:

    Editing must use hard cuts only: No fades, dissolves, or soft transitions.

    line 152
  22. Common:

    Aesthetic & Color Grading: Every cross-shot generation carries the same aesthetic header prompt: grain level, LUT direction, and light …

    line 197
  23. Very specific:

    After the complete MV prompt script is locked, write it to a dedicated markdown file complete-mv-prompt.md.

    line 218

“Common” is measured: kept by at least 6 in 10 of the 7 edited versions of this file we could read. “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

Tool mapping (mmx-cli edition)

This skill is the mmx-cli adaptation of the MiniMax-H3 Hub skill of the same name: the creative methodology is unchanged, while Hub canvas tools are replaced by the following portable equivalents.

Hub toolmmx-cli / portable equivalent
hub_generate_imagemmx image generate --prompt "..." --out <path>
hub_generate_videommx video generate --prompt "..." --out <path> (Hailuo; see mmx video generate --help for duration/model flags)
hub_generate_audio / hub_synthesize_speechmmx speech synthesize --text "..." --out <path.mp3>
hub_generate_musicthe minimax-music-gen skill (MiniMax music API; mmx has no music command)
hub_video_editffmpeg assembly (concat / xfade / audio mix) — deterministic, offline
hub_analyse_mediaffprobe + mmx vision describe <file>
hub_image_searchuser-provided assets first; mmx search query "..." for public references
hub_canvas_get_node / hub_canvas_group_recent_outputsthe episode working directory: assets-manifest.md, outputs/, decisions.md ledger files
choice cardsnumbered options presented in chat, choice recorded in decisions.md

Prerequisites: npm install -g mmx-cli and mmx auth login --api-key sk-xxx (region auto-detected; mmx quota shows the Token Plan balance). ffmpeg/ffprobe must be on PATH for assembly and inspection.

Music Aesthetics MV

Purpose

Use this Skill when the user wants to create, revise, audit, or generate music-video prompts or emotional short-film prompts where music, lyrics, typography, references, rhythm, performance, and camera language must be designed together. The workflow adapts MV prompt rules into mmx-cli execution: key creative decisions are confirmed, locked prompts are written to episode ledger files, and media creation runs through the mmx CLI.

Do not use it for ordinary subtitle burn-in, generic video editing, non-music product ads, or simple single-image / single-clip requests without MV structure.

Portability rules

  • Do not assume third-party tools, shell scripts, browser plugins, or external generation APIs.
  • Write locked prompts and revisions to the episode ledger (complete-mv-prompt.md, decisions.md).
  • Execute character cards, scene cards, typography cards, video clips, BGM, and final assembly via the mmx CLI (image/video/speech) and ffmpeg.
  • Do not hardcode output paths; always use Hub-returned paths.
  • If the user says “skip confirmation / just do it”, apply it only to the current run unless the outer orchestrator has active session authorization.
  • If the user only wants a prompt, stop after delivery. If the user wants a finished MV, continue to generation and assembly after confirmation.

Core principles

  1. Understand the work before choosing a structure. Templates organize, but are not mandatory forms.
  2. Use the latest confirmed creative intent.
  3. Omit irrelevant fields instead of filling placeholders.
  4. Do not mechanically add performance, typography, transitions, camera moves, or character actions just because a preset includes them.
  5. On-screen text is a designed visual layer, not ordinary subtitles, unless the user explicitly requests subtitles.
  6. If the user uploads a real song or beat, it is the master music bed unless they request replacement.
  7. Natural multi-shot stitching for >15s videos: when duration exceeds the model single-generation limit, use “multi-shot storyboard breakdown + tail/head frame continuation + beat hard cuts + global master-audio alignment”. Every segment must preserve the same groove, tempo feel, aspect ratio, character logic, visual preset, lighting language, and typography motion rules.
  8. If the user wants a finished music-aesthetic MV without lyrics, generate and lock original lyrics first. The final MV must reference character, typography packaging, and scene cards together. The typography packaging card controls only text packaging style, font texture, graphic design, layout ratio, and motion language.

STEP 1: Pre-flight lock

Before writing the final production prompt, confirm the minimum foundation with concise recommended choices.

1.1 Video format

Offer fixed options:

Use caseAspect ratioResolution
TikTok / Reels / Shorts vertical9:161080×1920
YouTube / Bilibili horizontal16:91920×1080
Feed square1:11080×1080
Dense vertical ad test9:16720×1280
Cinematic widescreen MV21:92560×1080

The selected ratio must be reused across character cards, scene cards, typography references, shot prompts, video clips, and final assembly.

1.2 Target duration and multi-shot structure

Before prompt writing or video generation, confirm target MV duration. Never silently use the model default as user intent.

Offer a compact duration card:

  1. 10-second test: one or two shots to verify character, scene, typography, and audio style.
  2. 15-second hook: 2–4 short shots for a more complete chorus / hook / performance phrase.
  3. 30-second or longer complete MV (multi-shot stitching): because many models have a 15-second single-generation ceiling, automatically use a Multi-Shot stitching workflow. First generate or lock one continuous song / BGM, split 30 seconds into 4–8 dynamic 2–5 second shots, then stitch via beat sync and head/tail frame continuity.
  4. Custom duration: user provides duration; plan shot count and stitching based on the selected model’s single-clip limit.

If uploaded music exists, the confirmed duration is the music window length.

1.3 Music window

If uploaded or specified music is longer than the confirmed duration, lock a segment before writing prompts:

  1. Recommended window: assistant identifies the strongest chorus / hook / emotional turn and requests confirmation.
  2. User timestamps: user provides start and end seconds; verify both are inside the audio.
  3. Multi-variant: only for ad-hook testing or multiple creative directions.

If music is shorter than the target duration, use the full audio unless the user explicitly asks to stretch or extend it.

1.3.0 Lyrics lock and lyrics-first fallback

Before final video prompts or generation, determine lyric ownership:

  1. If the user provides lyrics, those are locked lyrics. Do not generate, add, rewrite, expand, translate, paraphrase, or replace them unless explicitly requested.
  2. If the user wants a finished music-aesthetic MV without lyrics, generate short original lyrics matching the music style, vocal mode, duration, emotion, and visual preset, then lock them.
  3. Locked lyrics are the only source text for rap/singing performance and visible typography.
  4. Split locked lyrics into each shot’s Rap line:, Vocal line:, Soft vocal line:, or Spoken line:.
  5. Every Typography: field must come from the same locked lyrics. During vocal performance, visible text must word-for-word match the performed phrase.
  6. In the final typography MV, the performer must rap or sing the locked lyrics with visible lip shapes, jaw motion, breath, facial accents, nods, and hand accents following phrasing and rhythm.

1.3.1 Default multi-shot stitching for >15s videos

For >15s MVs such as 30 seconds, the global multi-shot prompt script is the authoritative source:

  1. Lock complete Master Audio: lock one continuous song / BGM track with the full vocal and groove as the only audio baseline.
  2. Build a Shotlist Timeline: split 30 seconds into 4–8 short shots of 2–5 seconds, mapped precisely to lyric timestamps and beats such as snare, 808, and drop.
  3. Head/tail frame and scene continuity: if Shot B continues the same scene, use Shot A’s tail frame as Shot B’s head frame. If it is a hard cut, preserve the same character card, wardrobe, lighting prompt, and use same-direction camera motion or match-cut visual elements.
  4. Generation and edit assembly: the video agent generates each shot from the Shotlist. The editing agent aligns all clips to the global Master Audio beat grid, trimming, speed-ramping, and stitching for natural lip-sync, beat sync, and transitions.

STEP 2: Creative Contract

Before final prompts, create a concise creative contract:

  • Music genre, instrumentation, tempo/BPM feel, vocal mode, and emotional temperature.
  • Lyric source and locked lyrics.
  • Target duration and multi-shot breakdown, e.g. 30s split into 6 short shots.
  • Reference image roles: character card, scene card, typography packaging card.
  • Camera language, shot sizes, focus, beat-cut density, and transition logic.
  • Exclusions such as no fades, no glossy AI beauty face, no single-shot overstretching that causes deformation.

STEP 3: Reference roles and scene sampling

Assign each reference image one narrow job:

  • Typography reference card: controls only text packaging style, font texture, graphic design, layout ratio, and motion language. It must not contribute people or scenes.
  • Character reference card: controls character identity, facial aura, hairstyle, clothing silhouette, proportions, posture, and presence.
  • Scene reference card: controls scene style, spatial atmosphere, image texture, background depth, and lighting mood.

3.1 Close-up scene switching and multi-shot sampling

For Trap, Dark-pop, Cyber-grunge, and other fast-paced MVs, switch quickly among multiple close-up spaces:

  • Keep scene mood unified, but each shot’s space must differ clearly, such as local background, tunnel, wall, light strip, or reflective floor.
  • Scene switches must be triggered by musical impacts: bass hit, 808 drop, snare, vocal accent, or typography smash.
  • Shot switches are “visual samples” on the trap beat: hard cuts, jump cuts, scan-glitch cuts, flash cuts, or action match cuts.

STEP 4: Preset grammar, Dark-pop / Cyber-grunge and Trap example

Visual and editing language

  • Realistic high-fashion texture and film-magazine texture, referencing late-90s / early-00s indie magazines, photocopy paper, film scans, and zine collage.
  • Heavy grain, slight film jitter, halftone dots, rough print edges, scan offsets.
  • Editing must use hard cuts only. No fades, dissolves, or soft transitions.
  • Image and typography strongly respond to beat cues: hi-hat roll creates micro-shake / frame skips, snare triggers scale-up / hard cut / shoulder drop, 808 bass hit creates low-frequency compression / stretch / offset.

Typography packaging rules

  • Typography is a dynamic graphic subject in space, not a subtitle bar. It may sit in foreground, midground, background, or be occluded by shoulders / hands.
  • Text must never cover eyes or main facial expression; avoid the mouth during critical lip-sync.
  • With vocals, visible words must exactly match the performed lyrics. Each shot has only one main typography event.

STEP 5: Prompt structure template

A multi-shot script must be modular by shot, with accurate duration and audio mapping:

[Global Aesthetic & Character Lock]: (global character and aesthetic anchor prompt)

Shot 1 (0.0s - 3.5s)
Vocal Line: "..."
Typography: "..."
Visual & Action: ...
Camera & Motion: ...
Transition Out: hard cut to Shot 2 on bass hit / same-direction whip pan

Shot 2 (3.5s - 7.0s)
Vocal Line: "..."
Typography: "..."
Visual & Action: ...
Camera & Motion: ...
Transition Out: ...

STEP 6: BGM continuity and natural multi-shot stitching system

6.1 Global audio continuity

The entire MV must bind to one Master Audio track. During segmented video generation, do not use independent disconnected clip audio. All video clips are aligned to their corresponding timestamps on the Master Audio timeline during edit assembly.

6.2 Natural Stitching Protocol

For >15s stitching, enforce five continuity locks:

Vocal & Lip Continuity: Cut points must land on lyric pauses, breaths, snare, or drop. Do not hard cut inside an active vowel or lyric unless the next shot is an extreme close-up with continuous mouth shape.

Rhythm & Beat Matching: Cut points must hit the 1/4 or 1/8 beat grid. Use speed ramping to align head nods, hand gestures, and blinks to beats.

Aesthetic & Color Grading: Every cross-shot generation carries the same aesthetic header prompt: grain level, LUT direction, and light direction. Final assembly uses unified 35mm film grain and LUT to hide batch color differences.

Transition & Motion Continuity: For long-shot continuation, use the previous tail frame as the next head frame and prompt continuation of action. For hard scene switches, use kinetic continuity such as same-direction pan or hand-occlusion match cut.

Typography Motion Stitching: Text in the previous shot should shatter, sweep out, or smash-screen on a bass hit; next-shot text smashes in or unfolds on the accent, transferring visual momentum.

STEP 7: Checklist

Silently verify before delivery:

  • If target duration exceeds 15s, has Multi-Shot breakdown been used automatically?
  • Is one global Master Audio locked, and are cut points beat-aligned?
  • Do cuts avoid mid-lyric vocal breaks, with natural lip and breath continuity?
  • Do adjacent shots have clear stitching logic such as tail/head continuation, same-direction motion, match cut, or beat hard cut?
  • Is the edit style strictly hard-cut, with no fades or soft transitions?
  • Are character / scene / typography cards isolated without cross-contamination?
  • Is typography a spatial layer that never blocks eyes or main facial expression?
  • Are image quality, grain, color, and lighting consistent globally?

STEP 8: Canvas delivery

After the complete MV prompt script is locked, write it to a dedicated markdown file complete-mv-prompt.md. The file must contain the full multi-shot prompt script and stitching notes. Later revisions update the same file.

STEP 9: Final MV generation and assembly workflow

  1. Prompt lock and ledger write: confirm the full multi-shot prompt script with timestamps, lyric mapping, and transition logic is written to complete-mv-prompt.md.
  2. Parallel shot generation: the video agent generates each 2–5s shot from the Shotlist. For continuation shots, extract the prior shot’s tail frame as the next I2V head frame.
  3. Multi-track editing and stitching: the editing agent imports the global Master Audio, aligns all shot videos by timestamp, trims and speed-ramps on the beat grid, and checks seams with same-direction motion or flash-cut masking.
  4. Finishing: apply unified Movie LUT and 35mm film grain overlay to reduce AI plasticity and lock visual consistency.
  5. Delivery: output the seamless complete MV video file.
<!-- QUALITY_BASELINE_V1 -->

When to use(什么时候使用)

当用户需要 在明确输入、预算和交付约束后执行生成或写入操作 时加载本技能。先从请求中提取目标、输入、约束、交付格式和验收标准;描述摘要为:For musicians, video creators, and social-media editors producing AI music videos or emotional short films with lyric typography. Users provide music, lyrics, references, characters, typography direction, mood, or target platform. The Skill analyzes beat and vocal timing, separates character, scene, and text references, designs beat-reactive spatial typography, decomposes long works into connected shots, audits prompts, and routes generation for H3 or other video tools. It outputs MV concepts, shot prompts, lyric text plans, and stitching guidance. Best for stylized MVs and subtitle-driven music visuals, not ordinary caption cleanup, licensed IP copying, or fully manual post-production editing.。

Rules

  • 先读后写:先确认当前状态与真实能力,再执行会改变外部状态的动作。
  • 权限最小化:只使用完成当前步骤所需的文件、工具、账户与网络范围。
  • 证据优先:运行结果、资源 ID、版本、哈希或测试输出缺失时,明确标记为 NOT_VERIFIED。
  • 幂等优先:保留请求标识与阶段状态;结果不明确时先查询,不进行盲目重试。
  • 隐私安全:日志、示例、回执和错误信息不得包含 token、cookie、密钥或个人敏感数据。

Workflow

Step 1:澄清意图

确认本技能是否匹配目标;若只是相邻需求,交给更精确的技能。

Step 2:执行预检

校验输入、模型/工具能力、输出路径、预算上限和审批状态;任一关键条件未知时停止在只读阶段。

Step 3:形成计划

列出将调用的工具、会改变的对象、成功标准以及失败后的安全退出方式。

Step 4:执行动作

按一次批准执行并记录请求标识;模糊结果先查询而不是重提;每个外部调用均保留可关联的状态或回执。

Step 5:验证交付

验证产物存在性、格式、哈希/标识、成本状态和质量门禁,并把事实、推断和未验证项分开陈述。

Validation checklist

  • 技能触发条件与用户意图一致,没有把相邻任务误路由到本技能。
  • 输入、目标对象、版本和输出位置均已明确,且没有使用猜测值替代必填值。
  • 所有写入、付费、发布或不可逆动作都在用户授权范围内。
  • 结果已用独立检查验证;仅有“命令成功”或“文件存在”不算完整验收。
  • 输出包含实际证据、失败/跳过项、剩余风险和可执行的下一步。

Gotchas

  1. 把计划当结果:文档或提示词不等于真实执行;必须标明实际运行层级。
  2. 错误重试:超时或响应丢失可能已经产生远端状态,先查询再决定是否重试。
  3. 隐式扩大范围:批量、全量、发布、覆盖和付费不是普通读写的自然延伸。
  4. 版本漂移:引用外部资源时记录版本、tag 或提交;不要把可变分支当发布证据。
  5. 证据过期:缓存、旧截图和历史测试不能证明当前环境;在交付前刷新关键证据。

不适用与边界

付费、发布、覆盖、上传或外部写入必须使用当前任务的显式授权;不自动扩大次数和预算。 如果请求需要别的技能,不复制其正文;按技能名称进行交接,并保留当前任务上下文。

Progressive disclosure

  • 需要确定输入/输出、状态和授权点时,读取 references/workflow-contract.md。
  • 需要交付前自检时,读取 references/validation-checklist.md。
  • 遇到超时、部分成功或恢复场景时,读取 references/error-recovery.md。
  • 首次运行、拒绝越权和失败恢复分别参考 examples/happy-path.md、examples/boundary-refusal.md、examples/failure-recovery.md。

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.