# Output spec ## Keyframe | Property | Value | |---|---| | Model | `fal-ai/nano-banana-pro/edit`, `aspect_ratio` 9:16, `resolution` 2K, with 1 to 3 pet photos (and an optional room photo) | | Format | JPEG, 9:16, at least 720 × 1280 | | Naming | `keyframe__.jpg`, e.g. `keyframe_biscuit_barista.jpg` | | Use | The start frame of the clip. Approved beside the pet photos before any video | ## Clip | Property | Value | |---|---| | Model | `fal-ai/kling-video/v3/pro/image-to-video` (default), `fal-ai/veo3.1/fast/image-to-video` or `bytedance/seedance-2.5/image-to-video`. Draft: `bytedance/seedance-2.0/mini/image-to-video`, 480p, silent | | Size | 720 × 1280 (9:16) after the finish | | Duration | 8 s (7 s if the end is trimmed) | | Frame rate | As the model made it, 24 to 30 fps | | Sound | Natural room sound, AAC 160 kbps, 0.1 s fade in and 0.3 s fade out. No speech; no music unless the scenario is musical | | Finish | `scripts/finish-clip.mjs`: scale and crop to 720 × 1280, H.264 High, CRF 20, 6 Mbit/s cap, `+faststart`, optional caption | | File size | Under 8 MB | | Text | None from the model. An optional caption drawn locally (white box, black bold sans, 15% from the top, at most 56 characters) | | Naming | `pet___9x16_720p.mp4`, e.g. `pet_biscuit_barista_9x16_720p.mp4` | ## Poster still | Property | Value | |---|---| | Source | The finished clip in the last 2 s, when the pet looks into the camera (`finish-clip.mjs --poster`, default 6.5 s) | | Size and format | 720 × 1280 JPEG | | Use | Cover frame in the app, thumbnails, any player that shows a still first | ## Web preview (for a page, not for posting) `finish-clip.mjs --muted` gives a silent copy under 8 MB; check it with `check-clip.mjs --muted`. ## Checks before delivery Run `scripts/check-clip.mjs --file clip.mp4 --pet pet.jpg --keyframe keyframe.jpg --seconds 8`, then open the board and the contact sheet, and listen once with sound on. 1. **Same pet.** On the board (pet photo, keyframe, first, middle and last frames): same breed look, coat colour and pattern, markings on the same side, eye colour, ear shape and size. A pet that turns into a different-looking animal fails. 2. **Anatomy.** Four legs, one tail, natural paws in every frame of the sheet. Paws push, press and pat; they do not turn into hands. 3. **Nobody else.** No extra animals. No people, or only the planned hand and forearm; never a face. 4. **Welfare.** Calm and playful throughout. No fall, slip, fear, cowering, injury or being trapped. No hazard in frame (heat, steam, knives, flames, strobes) and no harmful food. 5. **Openly staged.** Handheld phone look at eye level. Never a doorbell, security-camera or baby-monitor look (no fisheye, high corner, timestamp, night vision or greyscale). 6. **One take.** No cuts, jumps or black frames (the script flags them). A real light change, such as a fridge door, can be passed with `--ignore` after you have looked. 7. **The beat.** The action reads at phone size, and the look into the camera lands in the last 2 s. 8. **No text.** No words, logos or readable screens from the model. The caption, if any, is spelled exactly as the user gave it. 9. **Safe zones.** The pet's face and the action sit outside the top 12% and bottom 20%, and away from the right edge. 10. **Sound.** Room sound you can hear (the script flags a nearly silent track), no speech or voiceover, no music unless planned, nothing that sounds like a real song. 11. **Format.** 720 × 1280, 8 s, H.264, `+faststart`, under 8 MB (the script checks all of these). If check 1, 3, 4 or 5 fails, never deliver the clip as finished. Follow the fail path in SKILL.md. ## Posting note for the user Post it with the platform's AI label switched on, and do not caption it as something that really happened.