{"slug": "the-burning-bridges-ai-prompt-broken-down-shot-by-shot", "title": "The Burning Bridges AI Prompt, Broken Down Shot by Shot", "summary": "A developer published a shot-by-shot breakdown of the prompt used to generate the \"Burning Bridges\" AI party-video trend, in which a user's selfie is turned into a ten-second vertical clip of them rapping in a crowded, chandelier-lit room. The prompt is organized into four ordered blocks — format and camera, lead/crowd/setting, five timed beats, and sound — with hard rules placed first to prevent face drift, duplicate faces, unwanted text and background bleed from the reference photo.", "body_md": "You attach a selfie to an AI video tool, type something like \"me at a fancy party rapping to Burning Bridges, friends cheering\", and wait. What comes back is close but wrong. The person in the middle looks like your cousin. Three people in the crowd have your face. Your mouth opens and shuts at random. Or the tool refuses with a content warning and you never see a frame.\n\nThe trend clip is short and very specific: about ten seconds, one lead, a chandelier-lit room with green neon on the walls, a crew hyping them up, and five beats people recognise. A loose prompt gets you a loose party. Below is the prompt I use for it, taken apart line by line, with the failure each line is there to stop.\n\nIt has four blocks, always in this order:\n\n```\n1. Format and camera        one line\n2. Lead, crowd, setting     a few lines\n3. Five timed beats         one line each\n4. Sound                    one line\n```\n\nThe rules that must never break go at the top. The beats come after, once the model knows who is in the room.\n\n```\nA vertical 9:16 party clip, exactly 10 seconds, photoreal live action\nin full colour, handheld phone camera close to the lead in the middle\nof a packed crowd. No text, no subtitles, no logos, no watermarks.\n```\n\n**\"exactly 10 seconds\"** matters because the beat timings below add up to ten. Set the same length in the tool if it has a setting. If the clip comes out shorter, the last beats get squeezed or dropped, and the last beat is the payoff.\n\n**\"photoreal live action in full colour\"** pins the look so the model doesn't choose a style for you.\n\n**\"No text, no subtitles, no logos, no watermarks\"** is there because party and music-video footage is full of captions, on-screen words and labels on bottles, and models happily reproduce them.\n\nThe camera phrase changes with the format. More on that in section 8.\n\n```\nThe lead is the person from reference image 1, standing in the centre\nof the frame: their face, hair, skin tone, build and the exact clothes\nthey wear in the photo, sharp and recognisable in every shot.\n```\n\n**\"reference image 1\"** ties the lead to the photo, using the way many tools number their reference slots. With a single upload, \"the reference photo\" works too. The list of features gives the model something concrete to keep, since it has no idea who you are.\n\n**\"in every shot\"** is the line against face drift. The clip has jump cuts, and every cut is a chance for the model to rebuild your face slightly differently. Small changes add up, and by the last beat you can be a stranger. \"Every shot\" makes likeness a rule for the whole clip.\n\n**The clothes** are a trade-off. Keeping the outfit from your photo gives viewers one more clue that it's you. Asking for a night-out outfit looks more like a party but leaves the face to carry the likeness alone. Pick one and say it explicitly.\n\nIf your photo was taken in a kitchen, add:\n\n```\nUse only the subject from the reference image and ignore its background.\n```\n\nOtherwise your fridge tends to turn up behind the chandeliers.\n\n```\nThe lead is surrounded by a tight crew of eight to ten friends in\nstreetwear and caps, packed shoulder to shoulder on both sides and\nbehind, all hyped: dancing, raising silver cups, pointing at the lead,\nlaughing and cheering. Every friend is a generic extra with their own\nface and nobody resembles a famous person.\n```\n\nThe model has exactly one face to work from, yours, and you're asking it for a room full of people. Left alone, it tends to reuse the face it has, so you get the lead plus a few siblings. **\"with their own face\"** tells it the reference belongs to one person only.\n\n**\"eight to ten\"** is a size the frame can hold; \"a crowd\" could mean four people or a stadium. **\"on both sides and behind\"** keeps the crew out from between the lead and the lens.\n\n**\"nobody resembles a famous person\"** has two jobs. The trend comes from a music video, and a crowded party scene can pull the model toward familiar-looking faces. That's odd to watch and a filter risk, since most tools block real people.\n\n```\nSetting: a dim private dining room at night, ornate crystal chandeliers\nglowing gold on the ceiling, picture frames outlined in green neon on\nthe walls, long tables with white cloths, warm amber light.\n```\n\nThe tempting shortcut is \"like the Burning Bridges music video\". Don't. Naming the song or the artist is one of the most common reasons these prompts get refused, and the model has no reliable picture of that video anyway. Five or six concrete nouns it can render are enough. People recognise the trend from the room, the beats and the sound together.\n\n```\n0-2s: close-up, the lead points finger guns at the camera, cup in the\nother hand, friends bouncing behind.\n2-4s: rests a hand on their collar, chin up, eyes half closed, mouthing\nevery word.\n4-5.5s: turns to the friends at their side and points at them.\n5.5-7.5s: the friends crowd in around the lead, pointing at them and\nhyping them up, a hand swings past the lens.\n7.5-10s: the friends fill the frame around the lead, the lead covers\ntheir face laughing, then waves at the camera with the cup while\neveryone around throws their hands up and cheers.\n```\n\n*The five beats with their timings: finger guns, hand on the collar, point at the crew, the crowd closes in, the laugh and the cheer. Image created with AI.*\n\n**How the model reads it.** Each line opens with a time range, so the model gets order and duration at once. A paragraph like \"they rap, point at their friends, then everyone cheers\" gives the order but not the pacing, and you tend to get a long, wandering middle and no ending.\n\n**One action per beat.** Two seconds holds one gesture. When a beat needs two, \"then\" makes them a sequence instead of both at once.\n\n**The timings aren't equal on purpose.** The third and fourth beats are short (1.5 and 2 seconds) because pointing is quick. The last gets 2.5 because the cheer is the payoff and the loop point.\n\nYou can't ask a video model to perform the track. Song titles and lyrics get prompts refused, and the model couldn't match the recording anyway. Two lines handle the mouth instead:\n\n```\nThe lead raps along the whole time: mouth moving in time with the beat,\nfull of swagger, playing to the lens.\n\nSound: an original hip-hop beat with a rap vocal that matches the lead's\nlips, the crowd cheering. Not any real or recognisable song.\n```\n\nAsking for a vocal is what gets the mouth moving in rhythm. Write \"no singing, no dialogue\" and the mouth has nothing to do, so it looks dubbed the moment you add music. \"Not any real or recognisable song\" stops it imitating a real recording.\n\nThen swap the generated audio for the Burning Bridges song in your editor, or pick the official sound when you post, and line its start up with the finger guns. The mouth is matched to the beat rather than to each word, and at phone size that reads as rapping along.\n\nThis scene trips filters in predictable ways: drinks and nightlife words, real names, song titles and lyrics, photos that look under 18 (beauty filters make adults look younger) and photos with several faces. What the prompt does about it:\n\nThe [blocked video guide](https://tryburningbridges.com/guides/burning-bridges-ai-video-blocked?utm_source=devto&utm_medium=article&utm_campaign=launch) covers errors that only look like blocks, such as \"no face found\".\n\nChanging the aspect ratio setting isn't enough. The shot wording changes with it.\n\n|  | 16:9 wide | 9:16 vertical | \n|---|---|---|\n| Camera | handheld camera at chest height | handheld phone camera close to the lead | \n| First beat | (no framing word) | starts as a close-up | \n| Last beat | wide group shot | the friends fill the frame around the lead | \n\nA wide group shot in a tall frame makes everyone tiny, so the vertical version keeps the crew packed close. And cropping a finished 16:9 clip to 9:16 keeps only about 32% of its width; the [vertical vs widescreen guide](https://tryburningbridges.com/guides/burning-bridges-ai-video-vertical-or-widescreen?utm_source=devto&utm_medium=article&utm_campaign=launch) shows what gets cut.\n\nNo prompt line rescues a bad reference. The model copies what it can see.\n\n*One usable reference and three that cause trouble: sunglasses and hard sun, a small face in a group, a dark filtered profile. Image created with AI.*\n\nOne face, turned toward the camera, eyes open, in soft window light, and sharp. Head and shoulders are enough. Skip sunglasses, hands over the face, beauty filters and group shots.\n\nFor a pet, the identity line becomes \"keep its breed, size, coat colour and markings exactly as in the photo\", and the beats switch to an animal's moves: a paw lifted at the lens, a head bobbing to the beat, and a shake of the head and ears for the finish.\n\n| What went wrong | Line to check | \n|---|---|\n| The lead doesn't look like you | The feature list and \"in every shot\"; then the photo | \n| The crowd has your face | \"with their own face\" | \n| An extra looks like a celebrity | \"nobody resembles a famous person\" | \n| Beats out of order, or no ending | Timed lines, adding up to the clip length | \n| The mouth doesn't move with the song | The rap-along line and the vocal in the sound line | \n| Refused by the filter | Drinks, names, lyrics; then the gentler fallback | \n| Your kitchen is in the club | \"ignore its background\" | \n| Crew and hands cut off on TikTok | Vertical wording, not a crop | \n| Captions or logos appear | The \"no text\" line | \n\n*Disclosure: I build [TryBurningBridges](https://tryburningbridges.com/prompt?utm_source=devto&utm_medium=article&utm_campaign=launch), which runs a version of this prompt on one photo and lays the song over the result: a 10-second clip, vertical 9:16 or wide 16:9. It's paid, with one-time packs ($4.99 for 1 video, $11.99 for 3, $29.99 for 10) and no subscription. The link goes to our prompt page, with versions for a person and a pet to copy into your own tool.*\n\nStuck on a beat that won't behave? Leave a comment with the line you used and what came back, and I'll suggest a change.", "url": "https://wpnews.pro/news/the-burning-bridges-ai-prompt-broken-down-shot-by-shot", "canonical_source": "https://dev.to/tryburningbridges/the-burning-bridges-ai-prompt-broken-down-shot-by-shot-33lb", "published_at": "2026-10-12 07:08:38+00:00", "updated_at": "2026-10-12 07:09:33.218578+00:00", "lang": "en", "topics": ["generative-ai", "ai-tools", "ai-products"], "entities": [], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/the-burning-bridges-ai-prompt-broken-down-shot-by-shot", "markdown": "https://wpnews.pro/news/the-burning-bridges-ai-prompt-broken-down-shot-by-shot.md", "text": "https://wpnews.pro/news/the-burning-bridges-ai-prompt-broken-down-shot-by-shot.txt", "jsonld": "https://wpnews.pro/news/the-burning-bridges-ai-prompt-broken-down-shot-by-shot.jsonld"}}