shopping-shorts-factory

name

shopping-shorts-factory

description

End-to-end Korean shopping/product shorts video factory. Takes a Korean script (.md with Hook/Body/CTA structure) or product info and produces a 9:16 vertical MP4 (~30s) by orchestrating ElevenLabs TTS, Higgsfield CLI (Nano Banana Pro images + Kling 3.0 i2v), and ffmpeg composition (tpad freeze, atempo, ASS subtitle burn-in with Pretendard). Includes built-in A/B testing variant generation (shared body + alternate hook). Use this skill whenever the user asks to create Korean YouTube Shorts, Instagram Reels, TikTok-style product/shopping/beauty/cosmetic ads from a script, especially Hera/AHC/cushion/skincare/coupang-partners content. Also trigger when user says "쇼츠 만들어줘", "30초 영상", "AB 테스트 영상", "헤라/제품 쇼츠", "비포애프터 쇼츠", "비교 시연 영상", or references existing /04-scripts/*.md script files.

Shopping Shorts Factory

Korean shopping/product shorts automation. Single 30-second vertical MP4 from a script in ~5 minutes, ~40–70 Higgsfield credits per video.

When to use

• User has a Korean shorts script with Hook/Body/CTA structure (typically in 04-scripts/*.md)
• User wants to create a product/cosmetic/shopping ad shorts (Hera, AHC, cushion, skincare, coupang-partners, etc.)
• User asks for A/B test variants of a shorts video
• User explicitly says "쇼츠 만들어줘", "30초 영상", "쇼츠 영상 자동 제작"

Do NOT use when:

• User wants long-form (>60s) video — use a different pipeline
• User wants English narration — voice ID below is Korean-specific
• Higgsfield credits are below 100 (single video minimum ~40)

Output spec (always)

Field	Value
Aspect	9:16 vertical
Resolution	720×1280
Codec	h264 (libx264) + aac
Framerate	24 fps
Duration	~28–34s (slightly variable depending on narration)
Subtitle	ASS burn-in with Pretendard ExtraBold/Bold
Audio voice	Brian (Korean Seoul, male, advertisement style)

Prerequisites (verify before running)

Run these checks first. If any fails, stop and ask the user:

# 1. Higgsfield CLI authenticated
higgsfield account status
# Expect: "<email> — <plan> plan, <credits> credits"

# 2. ElevenLabs API key
# Check env var first, fall back to asking user
echo "${ELEVENLABS_API_KEY:-MISSING}" | head -c 4
# If MISSING: ask user for key, set via: export ELEVENLABS_API_KEY=sk_...

# 3. ffmpeg available
which ffmpeg && which ffprobe

# 4. Pretendard font installed
ls ~/Library/Fonts/Pretendard-ExtraBold.otf
# If missing: brew install --cask font-pretendard

Default voice: ElevenLabs voice ID r2b2z8wPmZeh7CQksHSs (Brian — Korean Seoul accent, male, advertisement category). Override with VOICE_ID env var if needed.

API key: Never write the ElevenLabs key into any file in this skill. Pull from $ELEVENLABS_API_KEY env var, or ask the user inline and set it for the session only.

Workflow overview

Korean script (.md)
   │
   ├─ Parse Hook / Body / CTA segments
   │
   ├─ [parallel] ElevenLabs TTS  → N segment mp3s (Korean, Brian voice)
   ├─ [parallel] Higgsfield image (nano_banana_2) → N reference PNGs (9:16, 2K)
   │
   ├─ Higgsfield video (kling3_0) i2v → N raw mp4 clips
   │     duration chosen per segment audio length (clamp 3–15s)
   │
   └─ ffmpeg assemble
         ├─ Normalize each clip: scale to 720x1280, pad, setsar=1, 24fps, libx264
         ├─ Match audio to video per segment:
         │     atempo (speed) if audio > video
         │     tpad (silence pad) if video > audio
         │     OR tpad clone (freeze last frame) on video if audio >> video
         ├─ concat video clips and audio segments
         ├─ ASS subtitle burn-in (Pretendard font dir flag)
         └─ Final mux: video + aac audio + faststart

Step-by-step

1. Parse the script

A typical script in 04-scripts/위크XX-제품-포맷.md has three sections marked by emoji headers:

• 🔴 훅 (Hook) — 0–3s tagline
• 📦 본문 (Body) — 3–25s main content
• 📢 CTA — 25–30s call to action

Extract the raw narration text (strip italic stage directions and inline markdown). Split body into 1–2 segments if its narration would exceed 15s (Kling 3.0 max clip duration).

2. Plan segments

Decide on N (usually 3–4 segments):

• Hook (1 segment, target 4–5s video)
• Body (1 or 2 segments, target 8–15s each)
• CTA (1 segment, target 4–5s video)

For each segment, draft a visual concept (start image) and an i2v motion prompt. Use "VIRAL HOOK STYLE" keywords for hook clips (zoom punch, jump cut, dynamic reveal).

3. Generate narration audio (ElevenLabs)

Use scripts/tts_korean.sh <text> <output.mp3>. Important constraint: max 3 concurrent ElevenLabs requests. Generate in batches of 3 or sequentially. After all segments are done, measure each with ffprobe to get exact durations — these drive the video clip length decisions.

4. Generate reference images (Higgsfield Nano Banana Pro)

Use scripts/gen_image.sh <prompt> — returns the image job UUID. Submit all N images in parallel, then wait for all. Cost: 2 credits per image at 9:16 / 1K-2K resolution.

Prompt patterns that work well for K-beauty shopping shorts:

• "Vertical 9:16, K-beauty editorial, hyper-realistic 4K"
• "Korean female influencer age 25–28, soft daylight"
• "Luxury cosmetic product on marble, golden price tag floating"
• "K-beauty advertisement final scene, glowing arrow pointing down"

5. Generate i2v video clips (Higgsfield Kling 3.0)

Use scripts/gen_video.sh <prompt> <duration> <start_image_id>. Cost: 1.75 credits per second (e.g., 5s = 8.75, 8s = 14, 12s = 21).

Strategy for matching audio duration → clip duration:

• Audio 4–5s → clip 5s
• Audio 7–9s → clip 8s
• Audio 10–13s → clip 12s
• Audio 14–15s → clip 15s (Kling max)
• Audio > 15s → split into 2 segments OR use freeze-frame extension in ffmpeg

For hook clips, the prompt prefix "VIRAL HOOK STYLE: Fast aggressive camera zoom-punch, dynamic reveal motion, jump-cut energy, …" produces engaging openings.

6. Assemble with ffmpeg

Use scripts/assemble_shorts.sh <work_dir> after laying out the directory like:

work_dir/
  audio/
    01-hook.mp3
    02-body1.mp3
    02-body2.mp3   (optional)
    03-cta.mp3
  video/
    01-hook.mp4
    02-body1.mp4
    02-body2.mp4   (optional)
    03-cta.mp4
  subs.ass         (subtitle file using assets/subs-template.ass styles)

The script handles:

• Per-segment audio fit: atempo (if audio > video) or silence pad (if audio < video)
• tpad freeze-frame extension (if video must be longer than recorded)
• 720x1280 normalization + 24fps re-encode
• concat with concat demuxer
• ASS burn-in with fontsdir=~/Library/Fonts
• Final mux to final.mp4

7. Open and verify

open <work_dir>/final.mp4

Visually inspect: subtitle timing, audio sync, hook impact. If body narration sounds rushed (>1.25x atempo), regenerate body video at longer duration instead of speeding the audio.

A/B testing variant (optional)

When user asks for A/B test (e.g., "A안 vs D안", "훅 변형 비교"), build two videos that share the body + CTA but differ only in the hook segment:

1. Generate one set of body/CTA assets (audio + images + videos)
2. Generate two hook assets (A-hook, D-hook) — different audio narration + different reference image + different i2v
3. Run assemble_shorts.sh twice with different hook clips

This keeps A/B isolation clean (single variable changed) and saves ~50% credits vs full duplication.

See references/ab-testing.md for KPI tracking guidance.

Cost estimation

Scenario	Images	Videos	Total credits
Single 28s shorts (3 segments)	3 × 2 = 6	5s + 10s + 5s = 35	~41
Single 33s shorts (4 segments)	4 × 2 = 8	5+8+12+5 = 52.5	~60
A/B test (2 videos, shared body)	5 × 2 = 10	5×2 + 8 + 12 + 5 = 52.5+8.75 = 61.25	~71

ElevenLabs cost: negligible (<500 chars per video, ~1 cent on the Creator plan).

Common pitfalls and fixes

Symptom	Cause	Fix
Failed to find two consecutive MPEG audio frames	ElevenLabs returned JSON error (rate limit) inside .mp3	Inspect file with head -c 200, retry sequentially (max 3 concurrent)
Body audio rushed >1.25x	tried to fit 18s narration into 12s video	Use freeze-frame extension (tpad) or generate 2 body clips
Subtitles render as squares/missing	Pretendard font not found by libass	Pass fontsdir=~/Library/Fonts to ass filter
Hook narration starts too early	no silence pad before TTS	adelay=800|800,apad,atrim=0:5.04 to delay 0.8s
Video clip resolution mismatch	Kling outputs vary (716×1284 or 768×1344)	Always renormalize with scale=720:1280:force_original_aspect_ratio=decrease,pad=720:1280:(ow-iw)/2:(oh-ih)/2,setsar=1
Error: unknown flag: --wait-timeout	wrong CLI flag	Use higgsfield generate wait <id> --timeout 5m --interval 5s

Reference files

• references/elevenlabs.md — TTS API details, voice settings, error handling
• references/higgsfield.md — CLI usage, model selection, cost math, JSON output quirks
• references/ffmpeg-pipeline.md — Composition recipes (tpad, atempo, ASS, concat)
• references/ab-testing.md — Single-variable A/B test methodology and KPIs

Bundled scripts

• scripts/tts_korean.sh — ElevenLabs Korean TTS (Brian voice)
• scripts/gen_image.sh — Higgsfield Nano Banana Pro image (9:16)
• scripts/gen_video.sh — Higgsfield Kling 3.0 i2v
• scripts/wait_higgsfield.sh — Poll job to completion + extract URL
• scripts/assemble_shorts.sh — Full ffmpeg pipeline

Assets

• assets/subs-template.ass — Pretendard-styled ASS template (Hook/Body/Price/Stat/CTA/BrandTag styles)

Output and logging

• Place all intermediate files under a per-run work directory (e.g., output/<slug>/)
• Final MP4 file naming: <topic>-<format>.mp4 (e.g., 위크03-헤라3종비교-A안.mp4)
• After successful production, write a work log to /Users/macbook15-platform/2026/L_2026_os/logs/ per the project's CLAUDE.md convention (date, tool, project, summary, file list, credit cost)