# Prompt patterns The keyframe decides what the street, the set piece and the product look like. The motion prompt decides what moves in 5 seconds. The sound line decides what the viewer hears. Approve both keyframes before any video is paid for. ## Locks Written once in stunt.json and reused through tokens: - `locks.product` (`{{PRODUCT}}`): the product's shape, proportions, parts and colours, written from the photo, with no logo, text or brand name. The only field left as FILL. "a low-top running shoe with a rounded toe, a thick cushioned sole, lace eyelets and laces, in white with soft green side panels and a green sole edge". - `locks.scale` (`{{SCALE}}`): how big, against something in the street ("as wide as the billboard it leans out of"). From the stunt. - `locks.street` (`{{STREET}}`): a type of street, from the menu or the user's own words. - `locks.light` (`{{LIGHT}}`): one light for both shots. - `locks.colour` and `locks.setpiece` (`{{COLOUR}}`, `{{SETPIECE}}`): the plain billboard panel or tram livery. - `locks.look`: the phone-footage lens and noise line. Generate shot 1 first, approve it, and put its path in shot 1's `keyframe`. Shot 2 then uses it as reference 1 (and the photo as reference 2), so the street, building, light and product match. ## Keyframe prompt anatomy `build-requests.mjs` wraps your text as: "Vertical 9:16 photo: one frame from a handheld phone video of a staged CGI street stunt, filmed on a phone's normal wide lens. {keyframe_prompt} {reference sentences} {logo sentence} {locks.look} Photorealistic, true-to-life scale... No readable text anywhere... People are tiny... Nothing is damaged... The product and the action sit in the middle of the frame, clear of the top 12% and the bottom 20%." Write `keyframe_prompt` as: camera position and height → the street and the set piece → the product at the key moment of the stunt, with its scale → the tiny people → the light. - Wide: "Wide establishing shot from the pavement across the street, the phone held slightly tilted up. The street: {{STREET}}. Above it is {{SETPIECE}}. The product ({{PRODUCT}}), {{SCALE}}, has pushed halfway out of the billboard towards the street as if the billboard were a window ... The billboard is intact, nothing is broken and there is no debris. Tiny pedestrians far below ..." - Hero: "Low-angle hero shot from the pavement, the phone held near the ground and looking steeply up. The product ... towers over the frame, leaning out of the same blank billboard ... seen with wide-lens perspective so it feels huge ..." Image model: `fal-ai/nano-banana-pro/edit` (`image_urls`, `aspect_ratio` "9:16", `resolution` "2K", $0.15, about 1536 x 2752). Both video models want a 9:16 image of 720p or more. ## Motion prompt anatomy `build-requests.mjs` builds: "Handheld phone video, vertical 9:16, one continuous shot, no cuts, real-time speed, the phone sways slightly like a person filming. {motion_prompt} The product keeps exactly the shape, colours and parts it has in the first frame and does not change size. Pedestrians stay tiny, far away and calm, with no face readable. Nothing is damaged and nothing falls. Audio: {sound}. No music, no speech, no voices, no sirens. No text on screen." Write `motion_prompt` as: the one move of the product → what stays still → what the people do (calm) → the camera. Shot 1 moves the product a little further along its stunt. Shot 2 mostly holds: the camera tilts up and stops, the product settles. A hero shot that barely moves is the point: it is the best chance to keep the shape. Write `sound` as the real sounds of the place in the order they happen: ambience first, then the product's own sound (a low rumble, a whoosh, a hum, dripping water), always "a soft crowd murmur with no words". Kling also sends the negative prompt from `assets/generation-input.json`. Seedance has no negative prompt field, so the "No music, no speech ..." line carries it. ## Failure modes and fixes | You see | Cause | Fix | |---|---|---| | The product changes shape or colour, a second product appears | Too much movement asked of it, or a busy keyframe | Say less: "eases a little further", "hangs almost still"; re-roll the keyframe against the photo; use Kling | | It looks like a toy or a cut-out, with no contact shadow | The keyframe has no light match | Name the light in `locks.light`; ask for "true contact shadows and reflections of this street"; re-roll | | Brand-like text on a shopfront, bus or billboard, badges on cars, number plates | The street invites signs | "blank fascias", "plain unbadged cars", "no number plates" are in the template; re-roll the keyframe and look at every sign | | A garbled logo on the product | The photo had one | Leave the logo plain (default), or `--keep-logo yes` and check letter by letter | | The billboard cracks, debris falls, the road breaks | The burst read as violence | "as if the billboard were a window", "intact", "nothing breaks or falls"; shorten the take to before it happens | | Kling tears a ragged hole in the billboard as the product moves out, with green bits falling (sample run, shot 1) | Any outward movement of a product that is half inside a panel reads as breaking through | Hold the product still and move only the camera (push in, tilt up): "does not move at all", "no tear, no crack, no hole". `stunts.json` does this for `billboard-burst` | | Kling invents writing on a blank panel: a brand-like word at 1.5 s, then script-like marks from about 3.3 s (sample run, hero shot) | Models fill a blank billboard with an advert as a shot goes on | Say "the billboard face stays one flat, plain, unmarked panel for the whole shot, with no writing, no logo and no advert"; this delayed the writing but did not stop it, so trim the hero `out` to before the first mark (the sample used 3.2 s) | | Big people in the near half of the keyframe, or a yellow number plate and a badge on a near car (sample run, first keyframes) | The model draws a street with life in the foreground | Put the pedestrians on the far pavement at 4 to 6% of the frame height, and say "very light traffic, small distant cars, none near the camera". A tiny, unreadable plate on a distant car can still remain: look at it zoomed in | | People run, point in alarm or crowd the product | The scale read as a threat | "calm and curious", "stand and watch"; make them smaller; re-roll | | A face is readable, or Seedance refuses the frame | A person is too big or turned to camera | Shrink or remove the nearest person in the keyframe; or use Kling | | The hero shot is a different street from the wide shot | Shot 2 was made without the wide frame | Make shot 2 with the approved shot 1 as reference 1; check the building and light side by side | | The clip looks like CCTV, news or a dashcam | A high fixed angle, a timestamp, a banner | Phone height, handheld, normal lens, no overlays; never accept it | | Music, speech or a siren in the sound | Model default for a street | "No music, no speech, no voices, no sirens" is in the wrapper; re-roll once, else deliver muted | | The move finishes at 3 s and the rest drifts | Too little for 5 s | Trim `out` to the clean part (`--only` the good shots) and say the clip is shorter | ## Drafts and finals - **Keyframes:** `fal-ai/nano-banana-pro/edit`, $0.15 each. - **Drafts:** `bytedance/seedance-2.0/mini/image-to-video`, 480p, 5 s, silent, about $0.36 each. A different model from the final: judge whether the move reads and the product holds, not the physics or the sound. If Seedance refuses the frame (a face), skip the draft and say so. - **Final (default):** `fal-ai/kling-video/v3/pro/image-to-video`, `start_image_url`, `duration` "5", `generate_audio` true, `negative_prompt`; $0.84 each ($0.168/s). Returned 1072 x 1928 in the sample run; assembly scales it. - **Premium final:** `bytedance/seedance-2.5/image-to-video`, 720p, 5 s, with sound, about $2.37 each. ## Fail path At most 3 attempts per shot at each paid step. Change one thing each time and say which: 1. Keyframe: the lock (more exact product description), then the camera or crowd, then the light. 2. Draft: the move (less, slower), then the keyframe. 3. Final: the move wording, then the sound line, then the model (`--model seedance`). After three, deliver the best attempt trimmed to its clean part and say exactly what differs. Never deliver a clip that fails the product-shape, text and brand, damage and panic, faces or tag checks.