跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
视频与音频

clip-machine

Turn ONE long video into native TikTok + Reels + Shorts posts — the four stations: CUT (word-level transcript, best moments, 9:16 reframe), CAPTION (burned word captions + per-platform caption text), COVER (per-platform cover rules), POST (one Zernio request per clip). Use when the user hands a long video and wants short clips selected, produced, and published natively everywhere. Triggers: 'run the clip machine', 'turn this video into clips', 'repurpose this video', 'post clips to tiktok reels shorts'.

Enriquemarq1 · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add Enriquemarq1/zernio-library-skills --skill clip-machine

技能说明


name: clip-machine description: "Turn ONE long video into native TikTok + Reels + Shorts posts — the four stations: CUT (word-level transcript, best moments, 9:16 reframe), CAPTION (burned word captions + per-platform caption text), COVER (per-platform cover rules), POST (one Zernio request per clip). Use when the user hands a long video and wants short clips selected, produced, and published natively everywhere. Triggers: 'run the clip machine', 'turn this video into clips', 'repurpose this video', 'post clips to tiktok reels shorts'."

Clip Machine — one long video in, native posts out

Four stations. Claude runs the first three, Zernio ships the fourth. No timeline software ever opens.

LONG VIDEO → [1 CUT] → [2 CAPTION] → [3 COVER] → [4 POST] → native posts live

The math this kills: 1 video → 3 clips × 3 platforms = 9 manual uploads, every week, forever. With the machine it's 3 requests.

Starting from zero? Copy-paste reference/initiation-prompt.md to Claude Code in an empty folder — it sets up the whole system for you.

STATION 1 — CUT

  1. Transcribe with word timestamps. WhisperX is the reference model — open source, free, runs on your machine (pip install whisperx, needs ffmpeg): whisperx video.mp4 --model large-v3 --output_format json The JSON gives every word a start/end timestamp via forced alignment (GPU: add --device cuda; CPU works with --compute_type int8). Any word-level transcription service works too — the machine only needs words with millisecond timing. No ground truth, no clean cut.
  2. Pick the best moments with the selection framework in reference/best-moments-prompt.md. That file IS the prompt — paste it with the transcript. Core law: finish the point — a clip starts at the beginning of a thought and ends after the payoff, never mid-sentence.
  3. Cut on word boundaries. Start ~0.15s before the first word, end ~0.25s after the last (never past the next word's start). ffmpeg: ffmpeg -ss {start} -to {end} -i long.mp4 -c:v libx264 -crf 18 -c:a aac clip.mp4
  4. Reframe to 9:16 (1080×1920). Crop to the speaker (face-tracking via MediaPipe if available; a fixed crop works when the framing is static). Screen-share sources: stack the screen crop on top, the face cam below.

STATION 2 — CAPTION (there are TWO captions — that's the teach)

  1. Burned-in word captions on the video. Most people watch muted. Word-synced from the transcript timestamps (that's why Station 1 needs millisecond words), 3-4 words per line, high-contrast, inside the caption-safe band (keep out of the bottom ~20% and top ~12% — platform UI lives there).
  2. Caption TEXT per platform — never the same text three times. Copy-pasting one caption everywhere is the amateur tell. Claude writes all three from the transcript:
    • TikTok: conversational hook + 3-5 niche hashtags.
    • Reels: the first ~125 chars must stop the scroll (that's all that shows before "...more").
    • Shorts: barely displays text — put the value in the TITLE instead. Full rules per platform: reference/platform-cheat-sheet.md.

STATION 3 — COVER

The first impression is platform-specific — one cover does not fit all:

  • Reels: the grid shows a 1:1 CENTER crop of your cover. Title text must live inside the center square or it gets amputated.
  • TikTok: pick the cover frame + big readable cover text (profile grid is how binge-viewers navigate).
  • Shorts: pulls a frame — make frame 0 count (cover-as-first-frame: the cover IS frame 0, the video starts on frame 1; no 1-second intro card — it kills retention). Details and dimensions: reference/platform-cheat-sheet.md.

STATION 4 — POST (the one-request money shot)

One request per clip through Zernio — the video + the per-platform captions

  • the platform list. Zernio uploads NATIVE to each platform (platforms bury links and reward native uploads — that's the whole reason the treadmill exists).
POST https://api.zernio.com/api/v1/posts
{
  "title": "Clip title (YouTube uses this)",
  "content": "fallback caption",
  "mediaItems": [{ "type": "video", "url": "https://.../clip.mp4",
                   "filename": "clip.mp4", "mimeType": "video/mp4" }],
  "platforms": [
    { "platform": "tiktok",    "accountId": "...", "customContent": "TikTok caption + hashtags",
      "platformSpecificData": { "privacy": "PUBLIC_TO_EVERYONE" } },
    { "platform": "instagram", "accountId": "...", "customContent": "Reels caption, hook in line 1" },
    { "platform": "youtube",   "accountId": "...", "customContent": "Shorts description",
      "platformSpecificData": { "title": "Value-first title #Shorts", "visibility": "public",
                                 "categoryId": "27", "madeForKids": false } }
  ],
  "scheduledFor": "2026-08-20T18:30:00Z"
}

Media upload for local files: POST /v1/media/presign → PUT the file → use the returned publicUrl in mediaItems. The full publishing flow (approval gates, scheduling, verification) lives in the zernio-publish skill — this station rides it.

Hard-won gotchas:

  • Post platforms SERIALLY, never two multi-platform posts in parallel.
  • A timeout does NOT mean the post wasn't created — GET the post before retrying (blind retries create 409s or duplicates).
  • Space clips out: 1-2 per day beats 9 in one hour.

The loop

Clips point viewers to the long video. The long video feeds the machine. Every clip carries an end-card or caption line back to the source — the machine is a circle, not a pipe.