Hell Grind: Rebuilding Our Biggest Fight Scene from Scratch

2026-08-05 · case-study tutorial · Adil · recording, 18:39 · watch on YouTube

Topic: Video generation & shot direction (also: Prompting & AI assistants, Assets, references & consistency) · mode: Tutorial · tags: prompts shared, one creator's experience, fiction, series, time-bound offer

Adil rebuilds the samurai battle scene of Hell Grind, which Higgsfield open-sourced with every prompt, asset and iteration, using a Claude skill that writes prompts with asset tags, distances in meters, timings in seconds and rules against hallucination. Shot by shot he fakes an army with distinct classes of extras, thick fog and three depth layers, writes camera orbits and actions in raw numbers, and locks matching poses across a cut for bullet time. His rule is to read what broke and fix only that; against the video-game look he swaps bans for direct descriptions, since models ignore "not", and has the model edit with four hard cuts.

What is said

The parts of the video and what is said in each (★ — how much it matters to the video).

00:00:00 — Hell Grind, made in 14 days and premiered in Cannes, is open-sourced

★★ Adil says Hell Grind, which he calls the first AI feature film, was made by the team in 14 days and premiered in Cannes. — 00:00:00 · Hell Grind
★★★ Because everyone kept asking how Hell Grind was made, Higgsfield open-sourced the whole project: every prompt, every asset and every iteration, ready to recreate it. — 00:00:00 · Hell Grind

00:00:29 — The $1 million contest and its first two judges

★★ Adil says Higgsfield announced this week (week of 3 August 2026) one of the biggest AI contests in the world, with a $1 million prize pool. — 00:00:29 · Higgsfield Global Film Festival
★★★ Adil names the $1 million contest's first two judges: Pixar co-founder Ed Catmull, a five-time Oscar winner, and cinematographer Phedon Papamichael, a two-time Oscar and BAFTA nominee. — 00:00:29 · Higgsfield Global Film Festival
★★ Adil says Ed Catmull and Phedon Papamichael are only the first two judges of the $1 million contest and more names are coming. — 00:01:04 · Higgsfield Global Film Festival

00:01:29 — The battle scene being rebuilt, played in full

00:03:09 — The Claude skill that writes the prompts: four parts, setup and what to feed it

★★★ After days of rewriting the same shots, the Hell Grind team packed its workflow into one Claude skill that structures prompts in four parts: asset tags as simple anchors, spatial coordinates in meters, time-coded movements in seconds and strict rules against hallucination. — 00:03:09 · Claude, Hell Grind · also in: Direct AI fight scenes through controlled iteration: Blueprint the fight before motion
★★ Adil's setup of the Hell Grind prompt skill, free in the description, takes about 30 seconds: in Claude go to Customize, Skills, press the plus button and upload it. — 00:03:09 · Claude, Hell Grind · also in: Direct AI fight scenes through controlled iteration: Blueprint the fight before motion
★★★ To get detailed prompts from the Claude skill, Adil feeds it the script page of the scene, attaches the assets with their names so Claude writes tags such as @Roko and @Monster into the prompt, and describes each sequence as he pictures it. — 00:03:51 · Claude · also in: Direct AI fight scenes through controlled iteration: Blueprint the fight before motion
★★ Before touching a prompt Adil keeps every asset of the scene on Canvas, his workspace, so everything is in view and nothing gets lost. — 00:03:51 · Canvas · also in: Direct AI fight scenes through controlled iteration: Blueprint the fight before motion
★★★ For the battle scene Adil's team made an extra character sheet of the hero Roko to make him look wiped out and in deep pain. — 00:03:51 · character sheet · also in: Direct AI fight scenes through controlled iteration: Blueprint the fight before motion

00:04:08 — Assets on Canvas: Roko's extra character sheet, the monster, the army extras and a two-angle location

★★ The battle scene's other assets are the monster (a 4-meter creature with a skull face), the red crystal blade, an enemy army of extras and a snowy plain trapped inside a wall of fog. — 00:04:25 · also in: Direct AI fight scenes through controlled iteration: Blueprint the fight before motion
★★★ For the snowy plain Adil built two distinct angles, a front view with a single crimson tree and an empty back view; locking both sides matters for action scenes, so the model stops guessing what is behind the character and batches stop drifting. — 00:04:25 · also in: Direct AI fight scenes through controlled iteration: Blueprint the fight before motion

00:05:00 — Shot 1: the opening one-take in three iterations

★★ Shot 1 starts from a simple text description: the monster mid war cry, then the camera pulling back as an army rises around Roko, all in one take. — 00:05:23
★★ In the prompt, each character tag (@Roko, @Monster) carries a short physical summary tied to its reference image, so the model knows who is on screen. — 00:05:23 · also in: Direct AI fight scenes through controlled iteration: Diagnose and repair a complex shot
★★★ The prompt's environment and style block locks weather, fog density, ground details and lighting; Adil says keeping these rules active across every frame stops the world changing mid-shot. — 00:05:52 · also in: Direct AI fight scenes through controlled iteration: Diagnose and repair a complex shot
★★★ For the drone move Adil's prompt asks for an FPV drone flying backward in a straight line at three times the speed of a typical cinematic dolly pull, since models understand direct comparisons better than descriptions. — 00:05:52
★★ Adil pastes Claude's prompt into Cinema Studio, where his assets, already saved as elements, snap into place, and sets the model to Seedance 2.0 with a 15-second duration. — 00:06:15 · Cinema Studio, Elements, Seedance 2.0, Claude · also in: Direct AI fight scenes through controlled iteration: Diagnose and repair a complex shot
★★★ Adil always generates a batch of four at once: if one take looks weird, he immediately knows whether it is a random glitch or the prompt. — 00:06:38 · rebatching · also in: Direct AI fight scenes through controlled iteration: Diagnose and repair a complex shot
★★ In the first Seedance 2.0 batch of the one-take, the warriors' emergence looked like a broken motion glitch and the army looked like a crowd of identical clones. — 00:06:38 · Seedance 2.0 · also in: Direct AI fight scenes through controlled iteration: Diagnose and repair a complex shot
★★★ Adil's fix for clone armies: do not just ask for a crowd but build distinct character classes for the extras; his two, a heavy masked brute and a light scout, give silhouette variety and keep the same warriors across generations. — 00:07:15 · also in: Direct AI fight scenes through controlled iteration: Diagnose and repair a complex shot
★★ The first one-take's camera path was too flat: it flew at the same height and speed throughout, leaving the characters as tiny dots in the fog by the end. — 00:07:15 · also in: Direct AI fight scenes through controlled iteration: Diagnose and repair a complex shot
★★ For the second iteration Adil opens on an extreme close-up of the monster's face, gives Claude the changes in plain text and gets back the updated prompt. — 00:07:44 · Claude · also in: Direct AI fight scenes through controlled iteration: Diagnose and repair a complex shot
★★★ The monster has no right hand, only a bone blade growing from the elbow, yet even with "no held objects, no handle" in the prompt the model kept giving him a hand holding a sword. — 00:07:44 · Seedance 2.0 · also in: Direct AI fight scenes through controlled iteration: Diagnose and repair a complex shot
★★★ Adil says telling the model what not to do rarely works and negative prompting was the wrong play; the fix for the monster's arm was simply to ask the model to raise the arm so it actually sees it. — 00:08:28 · also in: Direct AI fight scenes through controlled iteration: Diagnose and repair a complex shot
★★ In the final iteration of shot 1 the last camera move pulls up into a wide heroic frame of Roko surrounded by the army instead of staying low on his legs; that take made the final cut. — 00:08:28
★★★ Adil says the real skill of the workflow is not writing one genius prompt but reading what broke and fixing only that. — 00:08:56 · also in: Direct AI fight scenes through controlled iteration: Diagnose and repair a complex shot

“Not writing one genius prompt, but reading what broke and fixing only that.”
★★★ Adil's playbook for each shot: look at the generation, name what is broken in simple words and let Claude rebuild the prompt. — 00:09:24 · Claude · also in: Direct AI fight scenes through controlled iteration: Diagnose and repair a complex shot
★★ Adil says scale is where AI filmmaking differs from traditional production: an army shot would mean a CGI crowd simulation, months of work, days of rendering and a studio-sized invoice, but here it takes five minutes and one prompt. — 00:09:24 · also in: Direct AI fight scenes through controlled iteration: Direct scale and transformation with measurements

00:09:57 — Shot 2: faking an army of 1,000

★★★ Adil put the army scene in thick fog capping visibility at 20 meters on purpose: at that scale the model glitches in the background and the fog hides the mistakes, the cheapest cleanup there is. — 00:10:07 · also in: Direct AI fight scenes through controlled iteration: Direct scale and transformation with measurements
★★ Adil says fog matters less now that Seedance renders in 4K and holds the background much better, but the trick still works for packed scenes. — 00:10:07 · Seedance 2.0 · also in: Direct AI fight scenes through controlled iteration: Direct scale and transformation with measurements
★★★ The army prompt never says "show a thousand samurai": it builds the frame in three layers (one sharp warrior up front, a dense crowd behind, silhouettes fading into mist), so only about 40 soldiers are clear and the brain fills in the rest. — 00:10:57 · also in: Direct AI fight scenes through controlled iteration: Direct scale and transformation with measurements
★★★ Adil writes action second by second: the foreground samurai's rise goes in half-second steps (hand at 0.5, helmet at 1.5, standing by 3), and with the timing locked the model cannot rush or skip stages. — 00:10:57 · also in: Direct AI fight scenes through controlled iteration: Direct scale and transformation with measurements
★ The army shot ends with the whole army dropping into stance at once, eyes glowing red, and everything freezing, which Adil says gives tension and a free cliffhanger for the edit. — 00:11:23 · also in: Direct AI fight scenes through controlled iteration: Direct scale and transformation with measurements
★★ Once the army prompt clicked, the team ran a big batch, got several great takes and used two different clips from that one batch in the final edit. — 00:11:23 · rebatching · also in: Direct AI fight scenes through controlled iteration: Direct scale and transformation with measurements
★★★ For an arc around Roko, Adil writes the camera orbit in raw numbers (8 meters down to 4, 3 meters high down to 2), because video models prefer arithmetic over vague words. — 00:12:14 · also in: Direct AI fight scenes through controlled iteration: Direct scale and transformation with measurements
★★★ Roko's transformation gets a hard number, 0.4 seconds, and the prompt says it is a weapon deploying, not a magic sequence: a transformation left to take its time becomes a glowing anime power-up every time. — 00:12:14 · also in: Direct AI fight scenes through controlled iteration: Direct scale and transformation with measurements

00:12:30 — Shot 3: the transformation inside a camera arc

★ For the transformation shot's final edit the team pulled the best pieces from a couple of generations and mixed them together. — 00:12:42 · also in: Direct AI fight scenes through controlled iteration: Direct scale and transformation with measurements
★★ The first strike, Adil's favourite shot of the battle: a lead charger sprinting in with an odachi, a frozen moment before the weapons collide, a bullet-time orbit around the impact, then time snapping back as the blades shatter. — 00:13:33 · also in: Direct AI fight scenes through controlled iteration: Preserve physics through the impact cut

00:13:51 — Shot 4: the bullet-time first strike

★★★ Adil gives the model a strict match lock: the poses at the end of shot 1 exactly match the start of shot 2; otherwise the model treats the cut as a new scene and throws the poses away. — 00:14:00 · multi-shot prompt · also in: Direct AI fight scenes through controlled iteration: Preserve physics through the impact cut
★★★ For the slowed-down tension Adil maps the distance between the blades in centimeters, second by second, instead of writing "time slows down"; the Claude skill calculates the numbers. — 00:14:00 · Claude · also in: Direct AI fight scenes through controlled iteration: Preserve physics through the impact cut
★★ Adil writes the impact as physics: skip "sword clash", map the angles (a vertical chop meeting a diagonal parry) and set one hard rule, "crystal always beats steel", so the steel shatters. — 00:14:00 · also in: Direct AI fight scenes through controlled iteration: Preserve physics through the impact cut
★★ Adil calls the bullet-time strike one of the hardest shots in the film: it forces the model to track physics, two moving bodies, time dilation and a 180-degree camera arc at once. — 00:14:00 · also in: Direct AI fight scenes through controlled iteration: Preserve physics through the impact cut
★★★ The first bullet-time take was decent and its movement right, so Adil treated it as a batching game, not a prompt issue, and reran until he got the take; knowing a broken prompt from a bad roll saves credits. — 00:14:49 · rebatching · also in: Direct AI fight scenes through controlled iteration: Preserve physics through the impact cut
★★ The bullet-time shot was generated a few months earlier, when video models barely handled it; Adil expects that with today's updated models such a shot should land much faster. (hedged) — 00:14:49 · also in: Direct AI fight scenes through controlled iteration: Preserve physics through the impact cut
★★ When a shot keeps breaking, Adil suggests posting the image and the problem in Higgsfield's Discord, where others help tweak the prompt; he says half the tricks in the video came from that server. — 00:15:18 · Discord

00:15:55 — The final edit: the negation trap, direct descriptions and four hard cuts; the contest call

★★★ The full fight came out like a video game cutscene, the team's biggest struggle in every action scene of Hell Grind: writing "not a game, no CGI" a dozen times failed, because models ignore the word "not" and just see "game". — 00:16:17 · Hell Grind · also in: Direct AI fight scenes through controlled iteration: Replace game-like coverage with cinematic direction
★★ Adil says a continuous camera floating right behind the hero is straight out of a third-person video game, and real movies do not shoot like that. — 00:16:17 · also in: Direct AI fight scenes through controlled iteration: Replace game-like coverage with cinematic direction
★★★ Adil's first fix for the video-game look: swap bans for direct descriptions, since the model gives what you ask for, not what you ban; describe the lighting and textures you want instead of "no cheap effects". — 00:16:42 · also in: Direct AI fight scenes through controlled iteration: Replace game-like coverage with cinematic direction
★★★ Adil's second fix: since one long unbroken take is how video games look and real movies cut, the prompt makes the model edit with four hard cuts, each its own setup: a 50mm orbit, a 24mm low angle, an 85mm close-up and a 35mm wide. — 00:16:42 · multi-shot prompt · also in: Direct AI fight scenes through controlled iteration: Replace game-like coverage with cinematic direction
★★ Closing the rebuild, Adil says the $1 million contest is live now and asks viewers what they are going to build. — 00:17:45 · Higgsfield Global Film Festival
★★ With the four cuts the same fight looks like cinema to Adil: the model gave real cuts to edit with, and the different camera angles bring dynamic energy. — 00:17:45 · also in: Direct AI fight scenes through controlled iteration: Replace game-like coverage with cinematic direction
★ Adil calls the video lesson one and says the skill, the prompts and the whole playbook are free in the description. — 00:17:45

Threads in this video

milestone · Hell Grind (Higgsfield Original Series) — Because everyone kept asking how Hell Grind was made, Higgsfield open-sourced the whole project: every prompt, every asset and every iteration, ready to recreate it. (00:00:00)
milestone · Higgsfield Global Film Festival (2026) — Adil names the $1 million contest's first two judges: Pixar co-founder Ed Catmull, a five-time Oscar winner, and cinematographer Phedon Papamichael, a two-time Oscar and BAFTA nominee. (00:00:29)
reminder · Higgsfield: enter the Higgsfield Global Film Festival, submissions 10 August to 3 September 2026 (2026) — Closing the rebuild, Adil says the $1 million contest is live now and asks viewers what they are going to build. (00:17:45)
refine · How do you keep a character the same across shots? — For the battle scene Adil's team made an extra character sheet of the hero Roko to make him look wiped out and in deep pain. (00:03:51)
refine · How do you keep a location, product or prop the same across shots? — For the snowy plain Adil built two distinct angles, a front view with a single crimson tree and an empty back view; locking both sides matters for action scenes, so the model stops guessing what is behind the character and batches stop drifting. (00:04:25)
restate · How do you keep a location, product or prop the same across shots? — The prompt's environment and style block locks weather, fog density, ground details and lighting; Adil says keeping these rules active across every frame stops the world changing mid-shot. (00:05:52)
restate · How do you prompt a sequence of camera moves in one shot? — For the drone move Adil's prompt asks for an FPV drone flying backward in a straight line at three times the speed of a typical cinematic dolly pull, since models understand direct comparisons better than descriptions. (00:05:52)
refine · How do you prompt a sequence of camera moves in one shot? — For an arc around Roko, Adil writes the camera orbit in raw numbers (8 meters down to 4, 3 meters high down to 2), because video models prefer arithmetic over vague words. (00:12:14)
refine · When a generation fails, should you generate more takes, fix the prompt or start over? — Adil always generates a batch of four at once: if one take looks weird, he immediately knows whether it is a random glitch or the prompt. (00:06:38)
restate · When a generation fails, should you generate more takes, fix the prompt or start over? — Adil says the real skill of the workflow is not writing one genius prompt but reading what broke and fixing only that. (00:08:56)
first on record · How do you make a large crowd or army in a generated shot? — Adil's fix for clone armies: do not just ask for a crowd but build distinct character classes for the extras; his two, a heavy masked brute and a light scout, give silhouette variety and keep the same warriors across generations. (00:07:15)
refine · How do you make a large crowd or army in a generated shot? — The army prompt never says "show a thousand samurai": it builds the frame in three layers (one sharp warrior up front, a dense crowd behind, silhouettes fading into mist), so only about 40 soldiers are clear and the brain fills in the rest. (00:10:57)
restate · Should a prompt list what to avoid? — Adil says telling the model what not to do rarely works and negative prompting was the wrong play; the fix for the monster's arm was simply to ask the model to raise the arm so it actually sees it. (00:08:28)
refine · Should a prompt list what to avoid? — Adil's first fix for the video-game look: swap bans for direct descriptions, since the model gives what you ask for, not what you ban; describe the lighting and textures you want instead of "no cheap effects". (00:16:42)
restate · Which LLM should be used to write video prompts? — Adil's playbook for each shot: look at the generation, name what is broken in simple words and let Claude rebuild the prompt. (00:09:24)
refine · How do you avoid the AI (plastic) look? — Adil put the army scene in thick fog capping visibility at 20 meters on purpose: at that scale the model glitches in the background and the fog hides the mistakes, the cheapest cleanup there is. (00:10:07)
restate · How do you avoid the AI (plastic) look? — Roko's transformation gets a hard number, 0.4 seconds, and the prompt says it is a weapon deploying, not a magic sequence: a transformation left to take its time becomes a glowing anime power-up every time. (00:12:14)
refine · How do you get precise motion (dance, action, choreography)? — Adil writes action second by second: the foreground samurai's rise goes in half-second steps (hand at 0.5, helmet at 1.5, standing by 3), and with the timing locked the model cannot rush or skip stages. (00:10:57)
restate · How do you get precise motion (dance, action, choreography)? — For the slowed-down tension Adil maps the distance between the blades in centimeters, second by second, instead of writing "time slows down"; the Claude skill calculates the numbers. (00:14:00)
refine · How do you direct a multi-shot sequence? — Adil gives the model a strict match lock: the poses at the end of shot 1 exactly match the start of shot 2; otherwise the model treats the cut as a new scene and throws the poses away. (00:14:00)
restate · How do you direct a multi-shot sequence? — Adil's second fix: since one long unbroken take is how video games look and real movies cut, the prompt makes the model edit with four hard cuts, each its own setup: a 50mm orbit, a 24mm low angle, an 85mm close-up and a 35mm wide. (00:16:42)
refine · How do you spend fewer credits? — The first bullet-time take was decent and its movement right, so Adil treated it as a batching game, not a prompt issue, and reran until he got the take; knowing a broken prompt from a bad roll saves credits. (00:14:49)
refine · How do you get a cinematic look in a generated image or shot? — Adil's second fix: since one long unbroken take is how video games look and real movies cut, the prompt makes the model edit with four hard cuts, each its own setup: a 50mm orbit, a 24mm low angle, an 85mm close-up and a 35mm wide. (00:16:42)
progress update · Adil: leave the Hell Grind prompt adherence skill in the description (2026) — Adil's setup of the Hell Grind prompt skill, free in the description, takes about 30 seconds: in Claude go to Customize, Skills, press the plus button and upload it. (00:03:09)

Where the stories that moved here went next

Adil: leave the Hell Grind prompt adherence skill in the description (2026) — kept; last: progress update, 2026-08-05
Higgsfield Global Film Festival (2026) — running; last: milestone, 2026-10-10
Hell Grind (Higgsfield Original Series) — completed; last: looked back, 2026-10-20

Mentioned

Central: Claude · Hell Grind · Hell Grind prompt adherence skill · negative prompt · Roko · Seedance 2.0

Also: People: Ed Catmull · Phedon Papamichael

Projects: Higgsfield Global Film Festival

Organizations: Discord · Disney · Pixar

Concepts: AI look · bullet time (concept) · character sheet · multi-shot prompt · rebatching

Products: Canvas · Cinema Studio · Elements

Also centred on Hell Grind prompt adherence skill: How to Make Ultra Realistic AI Videos (28 Best Tips)

Also centred on negative prompt: Direct AI fight scenes through controlled iteration: Replace game-like coverage with cinematic direction

Also centred on Roko: Direct AI fight scenes through controlled iteration: Direct scale and transformation with measurements

Replies to

refers to · Hell Grind · World's First Ever AI Feature Film · Higgsfield Originals (2026)

← Previous: Hell Grind · World's First Ever AI Feature Film · Higgsfield Originals (2026) · Next → Best 5 Platforms to Access Seedance 2.5 in 2026