Azharthegreat/app-promo-video-studio

App Promo Video Studio

Make a finished, honest app promo video from emulator captures, then publish it to YouTube, your Facebook Page, Instagram and TikTok. Preserve the real app UI; never invent features, numbers or claims.

This skill is app-agnostic. It works for any app. Everything is spec-driven: a new app is a new JSON spec plus its own screenshots, demo records, CTA, disclosure lines and voice-over lines - no code changes. The Astra Phone Directory promo (10 Oct 2026) is the worked reference in examples/; copy it and swap the app's content. For each new app, confirm the one marketing promise, the real feature set, the safe demo data, the Play Store link, and the honest disclosure lines before producing. Publish to the brand's own YouTube channel, Facebook Page, Instagram and TikTok; only the per-app YouTube playlist differs between apps.

Reuse the loaded skill within a session. Discover tool paths at runtime (which ffmpeg, emulator -list-avds); do not hardcode machine paths, and keep in-repo paths relative.

Hard rules (do not violate)

• Only safe demo data in a public video. Fabricated demo records (random names/avatars, fake CNICs) are fine; NEVER a real private person's number or the owner's own. Confirm the demo set with the user if the app returns live third-party data.
• Show every data field. Each card scene displays the whole card; do not crop out a field.
• Honest copy. No invented features, numbers, awards or reviews. Keep the third-party-directory and number-removal disclosure in the CTA and captions.
• No em dashes anywhere (scenes, captions, titles, descriptions, commit messages). Use "-" or split the sentence.
• Custom thumbnail, always. Never let a platform auto-pick a frame. For a Reel / Short / TikTok (vertical) use the 9:16 thumb_9x16.png; only a landscape YouTube upload uses thumb_16x9.png. A YouTube Short center-crops a 16:9 thumbnail to a narrow ~9:16 strip, so keep the key content in the central ~405px column and verify by cropping the 9:16 centre before upload.
• Urdu captions/titles in ROMAN URDU (Latin), not Urdu script. The spoken VO stays Urdu; the written caption/title/description for the Urdu cut is Roman Urdu.
• AI disclosure: be accurate, never automatic. Read each platform's current definition of AI-generated, altered or synthetic content and answer it truthfully. The voice-over is a neural (synthetic) voice, so disclose it wherever the platform's rules cover synthetic audio.
• YouTube "Made for kids": always No. These promos are for adults.
• Draft first, publish only on an explicit per-platform yes. Publishing is outward-facing and irreversible; prepare everything private, then wait for the go. Tag the Play build and record each post URL + outcome afterward.

Workflow (six stages)

Full detail in references/pipeline.md. Scripts live in scripts/, an example spec in examples/ (add your own screenshots first).

1. Capture (emulator). Clean status bar with scripts/set_demo_statusbar.sh; screenshot with adb exec-out screencap -p (prefix /sdcard/ commands with MSYS_NO_PATHCONV=1). Prefer stills over screenrecord (it is choppy). Clear recent-searches so framing is consistent; crop each result card to the whole white card.
2. Compose scenes. python scripts/make_scenes.py <spec.json> <assets_dir> <out_dir> -> s1.png..sN.png. Grammar: hook -> home screenshot -> one card per demo record -> capability statement -> CTA. See references/pipeline.md for the design rules.
3. Voice-over (EN + UR). python scripts/gen_vo.py <vo_spec.json> <out_dir>. English en-US-AvaNeural, Urdu ur-PK-UzmaNeural. The script carries the Windows SelectorEventLoop fix and 403 retry. Write natural Urdu, not a literal translation.
4. Music. FFMPEG=<ffmpeg> scripts/build_music.sh <out.wav> <seconds> - continuous chord pad, loudnormed to -16 LUFS so it is audible under the voice. Pass a few seconds more than the scene total. Verify with volumedetect.
5. Assemble. python scripts/build_video.py --scenes <dir> --vo <dir> --music <wav> --out <mp4> --ffmpeg <ffmpeg>. Auto-times each scene to its VO line (so EN and UR both fit), crossfades scenes, mixes music under the voice. Verify duration, levels and frames.
6. Thumbnail + publish. python scripts/make_thumbnail.py <spec.json> <assets_dir> <out_dir> then follow references/publishing.md for the four channels (custom thumbnail, honest AI disclosure, draft-first, confirm the target channel before upload).

Quick start (from the example)

SK=<skills-dir>/app-promo-video-studio   # wherever the skill is installed
FF=$(which ffmpeg); W=<scratch_dir>
python "$SK/scripts/make_scenes.py"   "$SK/examples/video_spec.example.json" "$SK/examples/assets" "$W/scenes"
python "$SK/scripts/gen_vo.py"        "$SK/examples/vo_en.json" "$W/vo_en"
FFMPEG="$FF" bash "$SK/scripts/build_music.sh" "$W/music_en.wav" 33
python "$SK/scripts/build_video.py" --scenes "$W/scenes" --vo "$W/vo_en" --music "$W/music_en.wav" --out "$W/promo_en.mp4" --ffmpeg "$FF"
python "$SK/scripts/make_thumbnail.py" "$SK/examples/video_spec.example.json" "$SK/examples/assets" "$W/thumb"

Swap in Urdu by running gen_vo.py with examples/vo_ur.json into vo_ur, rebuilding music a few seconds longer (Urdu lines run longer), and assembling against vo_ur.

Make it robust and intelligent

• Auto-timing (build_video.py) sizes each scene to its voice line, so the same scenes render clean EN and UR cuts without hand-tuning offsets.
• Spec-driven (make_scenes.py, make_thumbnail.py read one JSON), so a new app is a new spec plus fresh screenshots, not new code.
• Verify every stage: capture by eye (right record, all fields), music volumedetect (~-15 dB mean = really in the mix), final render (duration + levels + frames from hook/card/statement/CTA).
• Fail honestly: if edge-tts is network-blocked, ship scenes + music + timecoded scripts and name the voice as the only blocked part. Never fake voice, data, or claims.

References

• references/pipeline.md - capture, scenes, VO, music, assembly in detail.
• references/publishing.md - the four channels, custom thumbnail, no-AI rule, accounts/profiles, per-platform checklist, draft-first gate.
• references/gotchas.md - every trap that cost time (silent music, masked music, edge-tts Windows fix, /sdcard path mangling, card framing drift, upload roots).
• examples/ - the reference spec and EN/UR voice-over lines; screenshots are not shipped.