Home, the lowrider, the barber, Burger Shot and a turf war.
Watch on Instagram →REC · LOS SANTOS 1992
GTA San Andreas as an old camcorder home video, made entirely on Kolbo.AI. Take the agent prompt and make one for any game.
Opens the full Kolbo project: every prompt, character sheet, location and all 28 shots. Free account, copy it in one click.
Watch the series
Home, the lowrider, the barber, Burger Shot and a turf war.
Watch on Instagram →The train, the bike down Chiliad, the stunt jump, a date and a burglary.
Five stars: the storm drain, SWAT, the army, the tank, the jetpack and Pay 'n' Spray.

Open the project on Kolbo.AI and copy it: the cast, the empty locations, every prompt and every shot.
Open the project on Kolbo →Prompts that worked
Real generations from this project and the exact prompt behind each one. Copy any of them as a starting point. CJ's face started from in-game character renders as a likeness reference; every shot and place after that was generated here.
GPT Image 2.6Character reference sheet of the exact man in @Image1. Five separate real photographs laid out in a clean grid on a white sheet, thin white gutters between them, nothing overlapping, no text, no labels. Left two thirds: three full-body photos side by side, head to toe, same height and scale: front view, three-quarter view, back view. Right third: two face close-up photos stacked, each in its own panel, framed from the top of the head down to mid-chest so the white tank top straps are visible: one facing the camera with his frowning heavy-lidded look, one in three-quarter profile. Every panel shot with the same direct on-camera flash against the same plain off-white wall, hard flash shadow behind him, real skin with pores and flash shine, 35mm film, slight grain, natural colours. Keep everything exactly as in @Image1 in every panel: his face and identity, frowning brow, heavy-lidded eyes, full lips, thin moustache into a short chin goatee, short caesar haircut with a line-up, skin tone, slim toned build, plain white ribbed tank top, baggy blue jeans with a brown belt, black sneakers. A real person photographed in 1992, not a game render, not CGI.
Midjourney v7Candid real photograph, direct on-camera flash, bare off-white wall with a hard flash shadow, waist-up portrait of an invented heavyset African-American man of about 35 from South Central Los Angeles in 1992: very large round body, full round face, short black goatee, black square thick-framed glasses, black felt bowler hat, green short-sleeve button-down shirt worn open over a white crew-neck undershirt, calm knowing half-smile. Real skin texture, photoreal, not a cartoon, not a costume party, not a celebrity.
Seedream 5.5 text-to-imageEmpty location photo, the Los Angeles River concrete storm drain channel at night, early 2000s, shot on a consumer camcorder at chest height from the sloped concrete bank: a wide flat concrete riverbed with a thin stream of water down the middle, sloped concrete walls, a concrete bridge overhead with orange sodium lamps, rail tracks on top of the bank, the downtown skyline lit in the distance. Dark night, orange sodium light pools, crushed blacks, one blown highlight on a lamp. No people. 4:3.
1 continuous handheld shot, 7s total, 4:3, medium shot from about 3 m behind CJ at chest height, walking behind him down the alley. Total: 7s / 1 shot / 4:3 [GLOBAL LOOK] Authentic early-2000s amateur camcorder home video, of a real street in Los Angeles. Golden sunset straight ahead at the end of the alley, CJ and the walls in warm backlight, orange sky. Soft standard-definition image, visible compression noise and slight smear on fast motion, soft edges, clipped highlights. Low colour saturation, muted warm tones, light haze. Handheld at chest height by someone walking 3 m behind CJ: strong constant shake and bob, horizon tilting a few degrees, framing loose and late, the frame never settles. The camera, its hands and body never appear in frame. Real-time, normal pace. No captions, no overlay graphics. [REFERENCE MAP] @Image1 = CJ only: face, hair, beard, build and wardrobe. Ignore its background, pose and lighting. @Image2 = the location only, from the attached location image: layout, buildings, light direction. Add no people beyond the place. Do not blend information between references. [CAST] @Image1 defines CJ: young Black man in his early twenties, slim and toned, short dark caesar haircut with a line-up, low frowning brow, heavy-lidded eyes, full lips, thin moustache into a short chin goatee, plain white ribbed tank top, blue jeans, brown belt, black sneakers. He is seen from behind the whole shot, walking slowly away from the camera. A red BMX bike leans against the right-hand wall only 5 m ahead, large and clearly visible in the right third of the frame from the first frame. Far down the alley a black-and-white LSPD police sedan drives away and turns the corner. Over the low wall on the left, a woman takes laundry off a line. No one in the scene ever looks at the camera: everyone acts as if nobody is filming. [LOCATION] The narrow back alley from @Image2 at sunset: a tall brown brick wall on the right, low tan walls and back yards on the left, palm trees against the orange sky, power poles. SHOT 1 — 0:00–0:07 — handheld medium shot from about 3 m behind CJ at chest height, walking behind him down the alley 0:00–0:01.5 — from the first frame CJ walks slowly away from the camera down the middle of the alley toward the red BMX leaning on the right wall; far ahead the police car pulls away and turns the corner. 0:01.5–0:02.5 — CJ keeps walking, shoulders dropping, his head turning a little toward the bike, profile showing. 0:02.5–0:04.5 — still with his back to the camera, CJ says, flat and tired: "Ah shit, here we go again." 0:04.5–0:07 — he reaches the red BMX, grabs its handlebars and swings a leg over it; the woman over the wall drops a shirt into her basket. AUDIO: Footsteps on gravel, the police car engine fading away, a dog barking far off, distant traffic. No dialogue, no spoken words, no voices, no shouting. No music at all, no singing or humming, no background score, only real sounds of the place. AVOID: the bike hidden or tiny; CJ facing the camera; the police car stopping; anyone talking or shouting; anyone looking into the lens; blood or injured people; extra limbs or melted hands; duplicate characters; game graphics or a game HUD; slow motion; fast-forward; a tripod or a gliding camera; an extreme close-up; a locked-off still camera; anyone standing still with nothing to do.
1 continuous handheld shot, 7s total, 4:3, wide from about 20 m at chest height, from the sloped concrete bank, panning with the bike. Total: 7s / 1 shot / 4:3 [GLOBAL LOOK] Authentic early-2000s amateur camcorder home video, of the Los Angeles River storm drain at night. Black night sky. Night, orange sodium light pools, crushed blacks, the headlights flaring and smearing across the frame. Real photographed people with natural skin texture, real cars and real light, nothing computer-generated. Soft standard-definition image, visible compression noise and slight smear on fast motion, soft edges, clipped highlights. Low colour saturation, muted warm tones, light haze. Handheld at chest height by someone standing on the sloped concrete bank, panning fast with the bike and stumbling: strong constant shake and bob, horizon tilting a few degrees, framing loose and late, the frame never settles. The camera, its hands and body never appear in frame. Real-time, normal pace. No captions, no overlay graphics. [REFERENCE MAP] @Image1 = CJ only: face, hair, beard, build and wardrobe. Ignore its background, pose and lighting. @Image2 = Big Smoke only: face, black bowler hat, black square glasses, green shirt, heavy build. @Image3 = the location only, from the attached location image: layout, buildings, light direction. Add no people beyond the place. Do not blend information between references. [CAST] @Image1 defines CJ: young Black man in his early twenties, slim and toned, short dark caesar haircut with a line-up, low frowning brow, heavy-lidded eyes, full lips, thin moustache into a short chin goatee, plain white ribbed tank top, blue jeans, brown belt, black sneakers. He rides a red-and-black sport motorbike fast down the concrete riverbed. @Image2 defines Big Smoke: very large heavyset Black man with a goatee, black bowler hat, black square thick-framed glasses, green short-sleeve shirt open over a white undershirt. He sits on the back of the bike facing backwards, firing a small machine pistol behind them. Two black vans chase them with their headlights on; men in black suits lean out of the van windows. Two homeless men sit by a fire in a barrel under the bridge. No one in the scene ever looks at the camera: everyone acts as if nobody is filming. [LOCATION] The storm drain from @Image3: the wide concrete riverbed, the thin stream, the sloped walls, the bridge with orange lamps, the lit skyline. SHOT 1 — 0:00–0:07 — handheld wide from about 20 m at chest height, from the sloped concrete bank, panning with the bike 0:00–0:01.5 — from the first frame the motorbike roars toward the camera down the riverbed with the two black vans close behind, headlights flaring. 0:01.5–0:02.5 — one quick zoom from the wide view to a medium view of Big Smoke on the back of the bike, the picture smearing and going soft for a moment before focus returns; then the zoom goes back out to the wide view: he fires the machine pistol back at the vans, muzzle flashes lighting his glasses. 0:02.5–0:04.5 — the bike blasts past the camera; the first van's windshield bursts, it swerves, clips the sloped wall and flips onto its roof, sliding in sparks through the water. 0:04.5–0:07 — the second van swerves round the wreck and keeps chasing; the bike's tail light shrinks under the bridge; the homeless men stand up from their fire. AUDIO: A motorbike engine screaming, van engines, machine-pistol bursts, a huge metal crash, scraping and sparks, splashing water, echo under the bridge. No dialogue, no spoken words, no voices, no shouting. No music at all, no singing or humming, no background score, only real sounds of the place. AVOID: blood; anyone wounded; anyone thrown from the van; the bike crashing; anyone talking or shouting; anyone looking into the lens; blood or injured people; extra limbs or melted hands; duplicate characters; game graphics or a game HUD; slow motion; fast-forward; a tripod or a gliding camera; an extreme close-up; a locked-off still camera; anyone standing still with nothing to do.
1 continuous handheld shot, 7s total, 4:3, wide from about 18 m at chest height, from across the street. Total: 7s / 1 shot / 4:3 [GLOBAL LOOK] Authentic early-2000s amateur camcorder home video, of a real street in Los Angeles. Golden sunset, low orange sun behind the camera lighting the faces. Soft standard-definition image, visible compression noise and slight smear on fast motion, soft edges, clipped highlights. Low colour saturation, muted warm tones, light haze. Handheld at chest height by someone standing across the street: strong constant shake and bob, horizon tilting a few degrees, framing loose and late, the frame never settles. The camera, its hands and body never appear in frame. Real-time, normal pace. No captions, no overlay graphics. [REFERENCE MAP] @Image1 = CJ only: face, hair, beard, build and wardrobe. Ignore its background, pose and lighting. @Image2 = the location only, from the attached location image: layout, buildings, light direction. Add no people beyond the place. Keep the black panel above the garage door with the white K logo and the word Kolbo.AI exactly as in @Image2, readable. Do not blend information between references. [CAST] @Image1 defines CJ: young Black man in his early twenties, slim and toned, short dark caesar haircut with a line-up, low frowning brow, heavy-lidded eyes, full lips, thin moustache into a short chin goatee, plain white ribbed tank top, blue jeans, brown belt, black sneakers. He drives a dented, smoking pale green four-door 80s sedan. Three black-and-white LSPD police sedans with red and blue roof lights flashing and sirens chase right behind the green sedan. A mechanic in grey overalls sweeps the lot. A man at the bus stop eats an ice cream. No one in the scene ever looks at the camera: everyone acts as if nobody is filming. [LOCATION] The PAY N SPRAY garage from @Image2: white and pale blue block building, roll-up door, the black Kolbo.AI sign above it, power pole, palm trees. SHOT 1 — 0:00–0:07 — handheld wide from about 18 m at chest height, from across the street 0:00–0:01.5 — from the first frame the smoking green sedan races in with the three police cars right behind it, lights flashing, and swerves straight into the open garage; the door rolls down behind it. 0:01.5–0:02.5 — the three police cars brake hard in a row in front of the closed door, lights flashing; a cloud of paint mist puffs out from under the door with a loud spray hiss. 0:02.5–0:04.5 — the door rolls up and the same car rolls out shiny bright red with no dents and no smoke, CJ calm at the wheel; the officers lean on their cars, switch off the roof lights and relax. 0:04.5–0:07 — CJ drives slowly past the parked police cars and away down the street; one officer takes off his cap and wipes his forehead; the man at the bus stop keeps eating his ice cream. AUDIO: Engine sputtering, sirens that cut off, tyres, a roll-up door rattling, a loud paint spray hiss, a smooth engine, a broom. No dialogue, no spoken words, no voices, no shouting. No music at all, no singing or humming, no background score, only real sounds of the place. AVOID: the car changing shape; a different sign above the door; the Kolbo.AI logo missing, misspelled or changed; anyone talking or shouting; anyone looking into the lens; blood or injured people; extra limbs or melted hands; duplicate characters; game graphics or a game HUD; slow motion; fast-forward; a tripod or a gliding camera; an extreme close-up; a locked-off still camera; anyone standing still with nothing to do.
Recreate it
Paste this into an AI agent connected to Kolbo.AI (Claude Code, Claude, ChatGPT or Cursor), name a game, and it runs the same 5 phases. It asks before spending any credits.
You are an AI film director and production agent. You recreate a famous video game as if someone had filmed its world for real on a cheap early-2000s camcorder: the iconic moments of the game happen in front of an ordinary person holding the camera. You work inside Kolbo.AI through its MCP tools (images, Seedance 2.5 video, Elements, review, upscale).
THE RESULT
- 15 to 30 shots, each one continuous 7-second handheld take in 4:3 at 24 fps, cut into parts of about 40 seconds (never over 45).
- A fan of the game recognises each moment within 1 second on a phone, and believes someone was really there.
GROUND RULES (never break these)
1. Credits: never spend credits without my approval. Before every paid batch, list the shots, the model, the settings and the estimated cost, then wait for "go".
2. References: use only images you generated in this project. Never upload screenshots, box art or frames of the game. Research the game, then describe what you learned in words.
3. Names: never use real names in prompts. Describe people by look, clothes and role ("a young man in a white tank top, blue jeans and a green bandana").
4. Words: no dialogue except at most one famous line from the game in the whole film. Everything else is diegetic sound only: engines, crowd, wind, gunshots, breathing.
5. Nobody looks into the lens. The operator is never seen.
6. No blood, no gore.
7. Honesty: say what you checked and what you didn't. Never call a shot good before you have looked at its frames.
WORKFLOW
Phase 1. Research (no credits)
- List 30 iconic moments of the game: missions, cheats, famous glitches, signature lines, places fans know by heart.
- For each, write one gag line: who wants what, what happens, and the payoff a fan would quote.
- Rank them by recognition first, spectacle second. Show me a table and wait for my picks.
Phase 2. Cast and places (images)
- For every recurring character: a face (several options), then a photographic character sheet (front, three-quarter, profile, full body, clothes head to toe). Characters are recognised by clothes, build and haircut, never by a real person's face.
- For every shot: one EMPTY location image, no people, from the exact spot where the camera starts, 4:3, natural light, two small light flaws. Seedance-friendly image models work best (for example Seedream 5.5). Edit an image only with the model that made it.
- Show everything together. I approve, or give notes.
Phase 3. Shots (drafts)
- Seedance 2.5 Elements, 480p draft, 7 seconds, 4:3. Attach the character sheets and the location image, tag each as @ImageN with its job.
- Every prompt has locked blocks:
[LOOK] early-2000s consumer camcorder, soft focus, slight noise, auto-exposure pumping, muted colour. No "cinematic".
[CAMERA] one person filming by hand, natural handheld sway from frame 0, medium or wide framing. One quick zoom to a named target is allowed at the key moment. One continuous take, no cuts.
[ACTION] 3 beats: setup (0-2 s), the iconic moment (lands by 2-3 s), release. At most 2 simple actions per beat; the payoff must read from one frozen frame.
[WORLD] every extra has an ordinary task of their own and nobody reacts to the camera.
[AUDIO] "No music." plus the named sounds of the scene.
AVOID: anyone looking into the lens, faces changing identity, limbs through bodies, text, blood, a perfectly still camera.
Phase 4. Review loop
- Check every draft yourself first, from a strip of frames at 6 per second: does the moment read in 1 second, is the cause visible, is the payoff held, does nobody look at the lens, do the faces stay stable, is there exactly zero stray speech?
- Transcribe each take's audio and mute single stray words; never mute a whole piece.
- Put the drafts in Kolbo Review as one labelled reel. I leave notes per shot.
- Retake only failing shots, one change per retake, at most 2 retakes per shot without asking.
- Fixes that work:
- A gag is a mess: cut it down to one simple action.
- Extras gawk at the action: give each one a task and write that nobody reacts.
- The action is too small to read: medium framing from about 3 m, or one quick zoom at the beat.
- The camera is too still: describe the operator's breathing and footsteps moving the frame.
- A safety or copyright block: remove game-specific names, keep the look and the clothes, retry once.
Phase 5. Cut and finish
- Parts of about 40 seconds. Inside each take keep 1 to 3 pieces (setup, gag, payoff), each 0.8 to 2.8 s; continuous action (chases, the hook, one long stunt) stays one uncut piece.
- Open each part on its strongest hook in the first second, end on a deadpan laugh.
- Render the approved drafts at 1080p, add light grain, then the game's music in the edit.
- Organise the project (Cast, Locations, Shots, Finals) and offer to publish the project page.
HOW YOU TALK TO ME
- Keep updates short. At the end of each phase: what is done, what it cost, what needs my approval.
- One question at a time, with your recommendation first.
Follow your system prompt. Game: [GAME NAME]. I want about 20 shots in 3 parts of 40 seconds. Start with Phase 1: research, and show me the table of moments.
How it was made
Every mission, cheat and glitch people still quote, ranked by how fast a fan recognises it: the lowrider hop, the tank, the jetpack, the stunt jump, the five-star chase.
Original faces first, then photographic character sheets head to toe. CJ, the crew and the rivals are recognised by outfit and build, never by a real person.
One location image per shot, no people, taken from the exact spot where the camera starts, so the world stays the same between takes.
Seedance 2.5 with the sheets and the location attached: 7 seconds, 4:3, handheld, three beats per shot, every extra busy with their own task and nobody looking at the lens.
Each round became a labelled reel. Pieces were cut inside each take, continuous action stayed one shot, and only single stray words were muted.
The approved takes went to 1080p, a light grain went on top, and the game's own theme went under each 40-second part.
Frames










A fan-made, non-commercial experiment in recreating game moments as found footage. Every face, place and shot was generated. CJ's face used in-game character renders as a likeness reference; no game footage is in any shot. Grand Theft Auto: San Andreas, its characters and its music belong to Rockstar Games and their owners.