# Prompt patterns The keyframe (Nano Banana Pro edit) decides who the pet is and where it is; the motion prompt only decides what happens in 8 seconds and what it sounds like. Approve the keyframe beside the pet photos before any video is paid for. ## The pet lock `pet.lock` in plan.json is one line, written from the photos, and it is said once at the start of every keyframe prompt. Name what makes this pet this pet, in this order: 1. Breed look, or "mixed breed" with the nearest look ("a scruffy terrier mix"). 2. Size, in a unit and against something ("about 35 cm at the shoulder", "a large cat, about 6 kg"). 3. Coat: length, texture, colours ("short, glossy black coat", "wiry tan and white coat"). 4. Markings and exactly where they are ("a white blaze down the nose", "a black patch over the left eye", "white socks on both front paws"). 5. Eyes and ears ("amber eyes, one ear up and one folded"). 6. Collar or harness, if it always wears one. Example: "a small scruffy terrier mix about 35 cm at the shoulder, wiry tan and white coat, a black patch over the left eye, amber eyes, one ear up and one folded, a red collar". Markings on the wrong side are the most common drift. Name the side, and check it on the board. If the photos show a mirror image (a selfie camera), say so and judge from the other photos. ## The pet photos - 1 to 3 photos of the user's own pet (or one they have permission to use). One clear side view and one face-on view are best; a third can show a marking. - Sharp, the whole pet in frame, daylight if possible. No people in them: crop them out first. - A room photo is optional: the user's own kitchen or living room, with no people and nothing private in view (post, screens, photos of family). ## Keyframe prompt anatomy `build-requests.mjs` wraps the clip's `scene` as: phone-video framing at the pet's eye level → the pet lock and "keep it exactly as it is" → the scene → the room line (if a room photo) → the people line → home light, phone noise, true anatomy → safe zones → no text, logos or readable screens. The full wrapper is in `assets/prompt-template.md`. Write `scene` as: where the pet is and what it is touching → the props of the job, unbranded → the light → where the camera is. Start the pose with `{{PET}}`; it becomes "The pet". Everything the action needs must already be in the frame (the jug, the dough, the laptop), because the video model should not invent objects. Good scene lines (from `assets/scenarios.json`): - Barista: "The pet stands with its front paws on the edge of a home kitchen counter, looking at a stainless steel milk jug filled to the brim. The jug stands beside a small unbranded espresso machine that is switched off..." - Video call: "The laptop is seen from the back, so its screen faces away from the camera and its cool glow lights the pet's face." Screens facing away keep faces and text out of the clip. - Fridge raid: "The fridge door stands open and its cool white light is the only light... Full-colour, slightly grainy phone video look, filmed by someone crouching in the kitchen doorway." ## Motion prompt anatomy `build-requests.mjs` wraps the clip's `action` and `sound` as: handheld phone, one take, eye level, no zoom, no cuts → the action → "keeps exactly its look" → "moves like a real dog/cat... calm and playful... never falls, slips or looks scared" → the people line → "Audio: ..." → "No music." (unless the scenario is musical) → "No speech, no voiceover, no subtitles or text on screen." Write `action` in three sentences that match the beats: "First ... Then ... At the end the pet ... looks into the camera ...". Use verbs paws can do: nudges, sniffs, pats, presses, tugs, kneads, tilts its head, nods, backs out. Write `sound` as the real sounds of that room and that action, in the order they happen. Name the materials: "the clink of the metal jug", "claws tapping on the worktop", "the hum of the fridge". The Veo and Kling finals also send the negative prompt from `assets/generation-input.json` (security camera, CCTV, doorbell camera, fisheye, timestamp, night vision, text, extra animals, extra people, human face, extra legs, deformed paws, morphing, cuts, slow motion, and music when the scenario has none). Seedance has no negative prompt field. ## Failure modes and fixes | You see | Cause | Fix | |---|---|---| | Markings move or swap sides, coat colour shifts | Weak lock, or one photo | Name the side in the lock; add a second photo; re-roll the keyframe | | The pet is too big or too small for the room | No size in the lock | Add a size against something ("its head just reaches the counter top") | | Paws grip, pour or type like hands | Action asks for hand work | Rewrite with push, press, pat, nudge; keep one action | | Extra paws, a fifth leg, a merged tail | Busy pose or a fast action | Simpler pose in the keyframe (sitting, both front paws visible); slower action | | A second animal or a person walks in | The camera reveals new space | Keep the camera still-ish; "No people and no other animals appear" is in the wrapper; re-roll | | Logos on the espresso machine, text on the fridge food | Branded objects invited | "unbranded", "plain", "no print" in the scene; re-roll the keyframe | | Faces or a call grid on the laptop | Screen faces the camera | Turn the laptop so its back faces the camera | | The clip looks like CCTV or a doorbell camera | High angle, wide lens, night look | Eye level, normal lens, full colour, "handheld phone"; never accept it | | The pet looks scared, cowers or slips | Spill or noise read as a threat | Make the spill slow and small, the look "sheepish"; re-roll; never deliver fear | | Speech or a music bed in the sound | Model default | "No speech, no voiceover" and "No music" are in the wrapper; re-roll once, else deliver muted | | The action ends at 5 s and the pet does something odd | Too little action for 8 s | Give the payoff a held beat ("and keeps looking"); or trim to 7 s in the finish | | The model refuses the keyframe | A face, a screen, a hazard or a hurt-looking pet in it | Remove it and regenerate the keyframe; or try Kling | ## Drafts and finals - **Keyframe:** `fal-ai/nano-banana-pro/edit`, $0.15, `aspect_ratio` "9:16", `resolution` "2K" (a 2K frame, well above the 720p the finals need). The alternative is `openai/gpt-image-2/edit` at `image_size` 864 × 1536, quality high, about $0.27. - **Draft:** `bytedance/seedance-2.0/mini/image-to-video`, 480p, 8 s, silent, about $0.58. It shows whether the action reads and the pet holds. It is a different model from the final, so do not judge its physics or sound. - **Alternative final:** `fal-ai/veo3.1/fast/image-to-video`, 720p, `duration` "8s", `aspect_ratio` "9:16", `generate_audio` true, about $1.20 ($0.15/s; $0.10/s silent with `--silent`). - **Final (default):** `fal-ai/kling-video/v3/pro/image-to-video`, `start_image_url` (not `image_url`), `duration` "8", `generate_audio` true, about $1.34 ($0.168/s with sound, $0.112/s silent). No aspect or resolution fields: the keyframe sets them, and it may return 1080 × 1920 (the finish scales it to 720 × 1280). - **Premium final:** `bytedance/seedance-2.5/image-to-video`, 720p, 8 s, with sound at the same price, about $3.78. ## Fail path At most 3 attempts per clip at each paid step. Change one thing each time and say which: 1. Keyframe: the lock (more exact markings or size), then the pose (simpler), then the props. 2. Draft: the action (simpler verbs), then the timing (hold the payoff longer), then the keyframe. 3. Final: the action wording, then the sound line, then the model (`--model kling`). After three, deliver the best attempt, trimmed to its clean part, and say exactly what still differs from the pet or the plan.