@acedatacloud/skills 2026.613.1 → 2026.614.1

package/AGENTS.md

@@ -26,6 +26,9 @@ Skills are located in the `skills/` directory (also mirrored to `.agents/skills/
 - **seedance-video** — Generate dance/motion videos with ByteDance Seedance
 - **wan-video** — Generate videos with Alibaba Wan
 
+### Video Production (workflow)
+- **marketing-video-factory** — Produce a short-form marketing/feature video end-to-end (orchestrates the media skills: generate assets → capture UI → assemble with FFmpeg → distribute)
+
 ### AI Chat & Tools
 - **ai-chat** — Unified LLM gateway — GPT, Claude, Gemini, Kimi, Grok (50+ models)
 - **google-search** — Search the web, images, news, maps, places, and videos via Google

package/README.md

@@ -42,6 +42,12 @@ Compatible with **30+ AI coding agents** via the [agentskills.io](https://agents
 | [hailuo-video](skills/hailuo-video/) | Generate videos with Hailuo / MiniMax |
 | [seedance-video](skills/seedance-video/) | Generate dance/motion videos with ByteDance Seedance |
 
+### Video Production (workflow)
+
+| Skill | Description |
+|-------|-------------|
+| [marketing-video-factory](skills/marketing-video-factory/) | Produce a short-form marketing/feature video end-to-end — generate assets + capture UI + assemble (FFmpeg) + distribute |
+
 ### AI Chat & Tools
 
 | Skill | Description |

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@acedatacloud/skills",
-  "version": "2026.613.1",
+  "version": "2026.614.1",
   "description": "Agent Skills for AceDataCloud AI services — music, image, video generation, LLM chat, web search. Compatible with Claude Code, GitHub Copilot, Gemini CLI, OpenAI Codex, and 30+ AI coding agents.",
   "keywords": [
     "agent-skills",

package/skills/marketing-video-factory/SKILL.md

@@ -0,0 +1,186 @@
+---
+name: marketing-video-factory
+description: Produce a short-form (9:16, ~30-60s) marketing / feature-explainer video end-to-end using AceDataCloud's own AI APIs — generate B-roll + hero images (Flux/Seedream/NanoBanana/Seedance), a music bed (Suno) and voiceover (TTS), capture real product UI with Playwright, assemble with FFmpeg (burned-in captions), upload to CDN, then distribute. Use when the user wants to make a promo/feature/explainer video, a TikTok/Reels/Shorts/Bilibili/Douyin short, "automated video marketing", or to turn a product feature into a vertical video.
+when_to_use: |
+  Trigger when the user wants to PRODUCE a finished short-form video
+  (not just call one model): a marketing/feature/promo/explainer short,
+  a vertical video for TikTok / YouTube Shorts / Reels / Bilibili /
+  Douyin / X / LinkedIn, or an "automated video factory". This skill is
+  the playbook that orchestrates the single-service skills
+  (flux-image, seedream-image, nano-banana-image, seedance-video,
+  veo-video, suno-music, fish-audio, short-url) into one pipeline.
+allowed_tools: [Bash]
+license: Apache-2.0
+metadata:
+  author: acedatacloud
+  version: "1.0"
+compatibility: Requires ACEDATACLOUD_API_TOKEN (model calls) and ACEDATACLOUD_PLATFORM_TOKEN (CDN upload) in .env (see _shared/authentication.md). Needs ffmpeg, python3 + pillow, and (for screenshots) playwright. Optionally pair with the per-service MCP servers (_shared/mcp-servers.md).
+---
+
+# Marketing Video Factory
+
+Turn a product feature into a platform-ready vertical short, made **with the same
+AI APIs you're advertising** (so the video is also a live demo of itself).
+
+> **Setup:** [authentication](../_shared/authentication.md) · async behavior:
+> [async task polling](../_shared/async-tasks.md) · tool-use: [MCP servers](../_shared/mcp-servers.md).
+
+## Strategy (read first)
+
+1. **Dogfood.** Generate every asset through `api.acedata.cloud` (Flux, Seedream,
+   NanoBanana, Seedance/Veo, Suno, TTS). The output *is* the proof.
+2. **Real UI is the spine; AI footage is the spice.** For a software/dev product,
+   show real screen-captures/screenshots of the product for the demo beats;
+   reserve AI-generated clips for intro / atmosphere / transitions. AI models
+   hallucinate fake buttons/code → never use them to depict the actual product.
+3. **Structure every video:** `Hook (0-3s) → Value (4-15s) → Demo (16-50s) →
+   CTA (last ~5s)`. One cut every 2-4s. Captions always on (≈85% watch muted).
+
+## Workflow
+
+1. **Brief** the feature: pull real value-prop copy from your docs (verbatim
+   on-screen text beats invented copy).
+2. **Script** a Scene-JSON: per scene = `{visual, duration, caption}`, following
+   the hook→value→demo→CTA skeleton.
+3. **Generate assets** (async — submit then poll): hero/B-roll images, B-roll
+   clips, a music bed, voiceover. Use `watermark=false`. Pin aspect to 9:16.
+4. **Capture UI** screenshots with Playwright (mobile 9:16 viewport).
+5. **Assemble** with FFmpeg: per-scene render → concat → mix music. Burn captions.
+6. **Upload** the final MP4 to the CDN; **distribute** per platform.
+
+## Recipe — generate an asset (submit + poll)
+
+```bash
+# Hero / B-roll image. Flux & NanoBanana honor true 9:16; Seedream size is an
+# ENUM (1K/2K/...) and tends to return 16:9 — use Flux for vertical heroes.
+curl -sS -X POST https://api.acedata.cloud/flux/images \
+  -H "Authorization: Bearer $ACEDATACLOUD_API_TOKEN" -H "Content-Type: application/json" \
+  -d '{"model":"flux-2-pro","prompt":"<scene visual>, deep violet/indigo, cinematic, no text","size":"9:16"}'
+
+# B-roll clip (Seedance is fast/cheap; Veo for hero quality). watermark off.
+curl -sS -X POST https://api.acedata.cloud/seedance/videos \
+  -H "Authorization: Bearer $ACEDATACLOUD_API_TOKEN" -H "Content-Type: application/json" \
+  -d '{"model":"doubao-seedance-2-0-fast-260128","content":[{"type":"text","text":"<motion>, 9:16, no text"}],"resolution":"720p","ratio":"9:16","duration":5,"watermark":false}'
+
+# Music bed — instrumental (see suno-music). Then poll /suno/tasks.
+curl -sS -X POST https://api.acedata.cloud/suno/audios \
+  -H "Authorization: Bearer $ACEDATACLOUD_API_TOKEN" -H "Content-Type: application/json" \
+  -d '{"action":"generate","prompt":"uplifting minimal electronic, premium tech","instrumental":true,"model":"chirp-v5-5"}'
+
+# Voiceover — OpenAI-compatible TTS, SYNCHRONOUS (returns audio bytes, no polling).
+# Generate ONE file per scene so the audio aligns to scene boundaries.
+curl -sS -X POST https://api.acedata.cloud/v1/audio/speech \
+  -H "Authorization: Bearer $ACEDATACLOUD_API_TOKEN" -H "Content-Type: application/json" \
+  -o scene1.mp3 \
+  -d '{"model":"tts-1-hd","input":"<scene narration>","voice":"nova"}'
+```
+
+All of the above return a `task_id` — **poll the matching `/<service>/tasks`** until
+`state`/`status` is terminal, then read the media URL (see _shared/async-tasks.md).
+The media is served from `*.cdn.acedata.cloud`. Per-model details: `flux-image`,
+`seedream-image`, `nano-banana-image`, `seedance-video`, `veo-video`, `suno-music`,
+`fish-audio` skills.
+
+**Voiceover (TTS):** `POST /v1/audio/speech` is the OpenAI-compatible route — it is
+**synchronous** (returns the audio bytes directly, no `task_id`/polling), models
+`tts-1-hd` (default) / `tts-1`, voices `alloy|echo|fable|onyx|nova|shimmer`, and it
+speaks both English and Chinese. (`/fish/tts` is an alternate voice-cloning route.)
+The endpoint does **not** return word timings — run WhisperX on the returned audio
+for karaoke, or distribute words across the scene duration proportionally.
+
+## Recipe — capture product UI (Playwright)
+
+```python
+# pip install playwright && python -m playwright install chromium
+from playwright.sync_api import sync_playwright
+with sync_playwright() as p:
+    b = p.chromium.launch(headless=True)
+    # 9:16 viewport so a vertical video doesn't center-crop a desktop layout.
+    ctx = b.new_context(viewport={"width":1080,"height":1920}, device_scale_factor=2,
+                        storage_state="auth.json")  # omit for public pages
+    pg = ctx.new_page(); pg.goto(URL, wait_until="networkidle"); pg.wait_for_timeout(3500)
+    pg.screenshot(path="shot.png")
+```
+
+**Public marketing pages capture anonymously; logged-in console/app pages need a
+saved `storage_state`** (a throwaway demo account) — without it they redirect to
+the login wall.
+
+## Recipe — assemble (per-scene render → concat → music)
+
+Do **NOT** build one giant filtergraph with all inputs — with several hi-res
+images + caption PNGs it OOMs ("Cannot allocate memory"). Render each scene
+to an intermediate, then concat (stream-copy), then mux music:
+
+```bash
+W=1080; H=1920; FPS=30
+SCALE="scale=$W:$H:force_original_aspect_ratio=increase,crop=$W:$H,setsar=1,fps=$FPS"
+
+# Per scene: a still (loop to DUR) or a clip (trim to DUR) + a transparent
+# caption PNG overlaid. Produce identical-codec segNN.mp4 (no audio).
+ffmpeg -y -loop 1 -t $DUR -i still.png  -loop 1 -i capNN.png \
+  -filter_complex "[0:v]$SCALE[bg];[bg][1:v]overlay=0:0:shortest=1[v]" \
+  -map "[v]" -an -t $DUR -c:v libx264 -pix_fmt yuv420p -r $FPS \
+  -video_track_timescale 15360 segNN.mp4
+# (for a clip scene: replace "-loop 1 -t $DUR -i still.png" with "-i clip.mp4")
+
+# Concat (all segments share codec/params) + mix music under, with a fade-out.
+printf "file 'seg00.mp4'\nfile 'seg01.mp4'\n..." > segs.txt
+ffmpeg -y -f concat -safe 0 -i segs.txt -c copy video_silent.mp4
+ffmpeg -y -i video_silent.mp4 -i music.mp3 \
+  -filter_complex "[1:a]volume=-3dB,afade=t=out:st=$(($TOTAL-1)):d=1[a]" \
+  -map 0:v -map "[a]" -c:v copy -c:a aac -shortest one-api-key.mp4
+```
+
+Captions: render each as a **transparent PNG** (Pillow) and overlay — portable
+across all ffmpeg builds (avoids the libfreetype `drawtext` dependency). Use a
+real bold sans (`C:/Windows/Fonts/arialbd.ttf`, `DejaVuSans-Bold.ttf`, etc.),
+3-7 words/line, white + black stroke; push to the **upper third during UI demos**
+so the product stays visible.
+
+> Reference implementation (Scene-JSON contract, caption-burn, render driver,
+> material-library convention): **AceDataCloud/PlatformStudio** — `app/contract.py`,
+> `app/pipeline/`, `scripts/render_veo_rough_cut.py`, `scripts/build_material_catalog.py`,
+> and the curated material index `materials/catalog.json` + `materials/curated.json`.
+
+## Recipe — upload to CDN + distribute
+
+```bash
+curl -sS -X POST https://platform.acedata.cloud/api/v1/files/ \
+  -H "Authorization: Bearer $ACEDATACLOUD_PLATFORM_TOKEN" \
+  -F "file=@one-api-key.mp4"   # -> {"file_url":"https://cdn.acedata.cloud/....mp4"}
+```
+
+Distribute with a short link (`short-url` skill) + per-platform metadata. Author
+once at 1080×1920 inside a centered 900×1400 safe box, then atomize to 16:9 / 1:1.
+
+## Creative cheat-sheet
+
+- **Length:** TikTok 15-34s · Shorts/Reels 30-60s · X <60s · LinkedIn 30-90s.
+- **Captions:** always on, karaoke word-highlight, avoid bottom ~20-25% & right edge.
+- **Audio:** final mix ≈ **−14 LUFS**, music **15-20 dB under** voice; sidechain-duck
+  (`sidechaincompress`) the music off the narration; `loudnorm=I=-14:TP=-1:LRA=11`.
+- **CTA:** one clear visual CTA in the last ~5s (assume muted).
+
+## Gotchas
+
+- **Generation is async** — every model call returns a `task_id`; poll the
+  service's `/tasks` endpoint. Don't treat the first response as the result.
+- **Aspect:** pin 9:16 explicitly. `seedream` `size` is an enum (1K/2K/3K/4K) and
+  skews 16:9 → use **Flux** (`size:"9:16"`) or **NanoBanana** (`aspect_ratio`) for
+  true vertical heroes.
+- **`watermark:false`** on marketing assets (Seedance/Seedream default to true).
+- **Single mega-filtergraph OOMs** on multi-input hi-res jobs → render per-scene
+  then `concat -c copy` (above).
+- **Screenshots:** use a **9:16 viewport** (not a desktop one center-cropped);
+  logged-in pages need a saved Playwright `storage_state`.
+- **CDN download SSL:** some hosts (e.g. `cdn1.suno.ai`) fail Python `urllib`
+  cert verification → fetch with `curl -L` instead.
+- **Music licensing:** Suno free tier is **non-commercial**; paid plans give no
+  indemnification → for **paid ads** use a fully-licensed library; keep Suno for
+  organic. Master to −14 LUFS regardless.
+- **Don't hard-couple the video model** — keep it swappable (Veo/Seedance/Luma);
+  abstract the provider so a deprecation (e.g. a Sora API sunset) is a 1-line change.
+- **Cross-posting:** strip watermarks before re-uploading; rewrite title/hashtags
+  per platform; keep Instagram reposts <10 / 30 days; use a humane cadence (bursts
+  read as spam).