# FREE Claude Prompt That Creates AI Videos AUTOMATICALLY!

Channel: Parker Prompts
Video: https://www.youtube.com/watch?v=7bE0PCee01Y
Duration: 13 min
Language: English
Words: 2659
Transcript page: https://viewrankai.com/tools/youtube-transcript/7bE0PCee01Y

---

[0:00] Claude is one of the best AI models out there and it still has no way of making an actual AI video. But I wrote a prompt that makes the whole thing possible and simple enough to run in one chat. So I'm giving that prompt away for free and by the end you'll be making AI videos automatically inside Claude without

[0:13] opening any separate window.

[0:20] It's a nice room. [music] Look at this beauty.

[0:39] Which way? Right.

[0:46] [music]

[0:55] For this whole workflow, we're using Claude with the Higsfield MCP connected to it. And the link for that is in the description. So, let's open the Higsfield MCP page. And as you can see, the connector URL is right there in the first box. I'll copy that, go into Claude, and open settings. Under connectors, we add a new one, call it Higsfield, and paste the URL straight in. Then, I sign in, and the setup is done. The connector is exactly the same in Claude desktop and on Cloud.ai, so it

[1:18] makes no difference which one you're in. And there are no API keys to manage anywhere in this. I'll open a fresh chat and paste the prompt in. The full prompt is in the free doc below as well, so you can copy it out and run it alongside me while you watch. It breaks the whole job into eight stages, something I'd call stage gating, and then it stops at every one of them for my approval. So, it asks for a brief first. With the brief answered, it comes back with pitches, and the chosen pitch turns into a breakdown and then reference images. And video is stage six of the eight with the

[1:43] upscale and the edit notes after it. Engaging it like that means every decision comes back to you in writing first, so you can change it before it costs you a single credit. Claude comes back saying the MCP is connected and responding so it can make images and video inside this conversation now. But the answers we give next decide how the entire film gets built. So Claude opens with nine questions split into two batches so it reads like a conversation instead of a form. Five plain ones come first. It asks what type of video this is, what it's about, and the genre and tone. Then the runtime and whether we

[2:11] want dialogue in it or just visuals. So I tell it we're making a short film that's also a heist and a comedy thriller and tone. 45 seconds and yes to dialogue. But the technical questions are different because those already have defaults in them. So you don't need to know which video model is good at what or what resolution to start at because the prompt has already picked sensible answers and just asks you to confirm them. GPT Image 2 handles every image at high quality in 4K and Cance 2.0 handles the video at 1080p. Cance 2.5 is out now and it doubles the cap to 30 seconds a

[2:39] generation. So the same 45 seconds would come down to two. I'm staying on 2.0 here because it's the prompt's default and everything you're about to see was built on it. But if 2.5 is in your model list, the maths just gets easier. I'll personally set the aspect ratio to 21x9 because that's the widescreen ratio films get shot in. That one's for the video and the reference sheets come out at 16x9 because none of them are in the finished film. And we generate all the video at 1080p and then upscale right at the end because native 4K costs far more per render. So a scene that comes back

[3:07] wrong at 4K cost you a lot more than the same scene coming back wrong at 1080p. Now with the models and the ratio locked, the runtime [music] is the last thing it has to work out. Cance 2.0 0 caps out at 15 seconds per generation. So, our 45 seconds splits into exactly three. That's three generations and not 30 separate shots. And it works because each one can hold its own hard cuts inside it. That's a bit of maths I'd have got wrong on a first attempt. And the prompt has already done it. And it

[3:31] locks the whole lot back in one line. Models resolution, ratio, and block count. [music] Then three pitches come straight back. Each one three or four sentences with a title on it. And the premises are properly different from each other. Wrong floor is an office job where they realize they're one story too high. and the long wait keeps the whole robbery offcreen and stays on the driver's face. I'll take the first one, the nice room. It gives us a safe cracker in her 60s and an enormous, completely useless young accomplice. And the house they're robbing starts sealing itself while they're both still inside it. And the pitch never says how those

[3:59] three renders are supposed to join up, which is the next thing the prompt sorts out. Turning three separate renders into one continuous take comes down to one rule about where a part is allowed to end. Now Claude splits the film into three parts, one for each generation, and every part except the last has to end midaction. And the next part opens as a match cut on that exact action from a different angle. That is the trick for making three separate generations play as one film. And the prompt writes every part to that rule for you. So the finished thing looks like someone shot it. Each of the three parts comes back

[4:27] with a name and the rooms it's locked to. The plant ends mid lift with the box half clear of the safe. The break runs the study into the corridor and ends mid-stride and the leap goes through the window and cuts to black on the landing. Dialogue gets written at this stage, too, and there are only five short lines in the entire film. I'll approve every one of them word for word before anything renders. That order matters because changing a line after the scene is generated means paying to generate it all over again. And the first image it

[4:52] generates is Marggo's reference sheet. And one of the three panels on it has no head. Margot and Dustin get built as images before anything moves. And so do all three rooms they run through. Claude writes both character sheet prompts out in full and shows them to me before it generates a single thing. Each sheet comes back as one image with three panels inside it. On the left is the full body from the front with no head, no neck, and no hair on it, which gives the video model the wardrobe on its own as an unbroken silhouette with no face in the shot. In the center, the same

[5:18] body from behind with the head on, and that covers the angles a chase actually needs. On the right, a tight chest up shot. That one is the identity lock for the whole film, but both sheets close on the same block of instructions at the bottom. That block sets an 18% neutral gray backdrop, which is the plain mid-grade that photographers shoot against with flat shadowless light. And there's a line telling it to relight from scratch. That flat light is what

[5:39] makes the sheet usable as a reference. Any lighting that's already in a reference image shows up in every generation after it, and then the scene comes back lit wrong. It checks each sheet against its own prompt panel by panel. So, I approve them one at a time. Then, it saves them as Margot and Dustin

[5:52] and moves straight on to the locations. It comes back wanting five sheets and the moonlight on the wrong side. So, I cut it back to four and put the light where the study needs it. One prop sheet for the jewel box first and then three location plates. The box gets its own sheet because it appears in all three generations and it opens on camera. So, it gets a six panel grid covering closed and open from the front, the rear, 3/4 and straight down. Every plate is shot wide, empty of people, and straight on at standing eye level. The geography gets fixed left to right, so the safe is

[6:19] always on the left, and the window is always on the right. None of the plates carry any readable text either, and that one is written into the prompt as a hard rule. The corridor gets three moonlit stripes across the runner, one for each window, put there for later. And the gallery gets a retracted shutter housing above the arched window, which explains why that one window is still open. So, let's approve them and move on. So, six elements are live in total. That's Margot, Dustin, and Jewelbox for the people in the prop. Then, study corridor and gallery for the rooms. So, every prompt from here calls a tag instead of

[6:47] describing anyone again. That tagging is what stops the faces changing between renders. But the obvious way to prompt a video is to write the shot you want as a paragraph, and this prompt never does that. When I wrote video prompts as one long paragraph, my two characters kept swapping sides of the frame. So, this prompt uses labeled blocks instead, and we're filling in the first of them right now. The blocks cover the scene context, which elements are active, and a map of the location, then the first frame, and the format, followed by action, timing, physics, lighting, and audio with a set of positive locks at the bottom. But the

[7:15] location map is the block that stops the swapping. It says Margot owns the left of the frame and Dustin owns the right in every shot of every generation, so they never swap sides between renders. By hand, you'd have to spell all of that out again for every shot. And the blocks mean you just fill in a short line under each heading instead. So, scene one runs six shots inside one 15-second generation with hard cuts between all of

[7:34] them. Touch nothing.

[7:44] It's a nice room. The very first frame already has both of them in position, so we don't waste any of the 15 seconds on an empty establishing shot. And it ends exactly where part one said it would. With the box half clear and her hands still rising, you state what happens and never what shouldn't. And the only sound allowed inside a generation is sound

[8:04] that's actually happening in the room. So no music and no score anywhere inside any of the three. The turn in this film is a single sound, the steel shutter slamming into the sill at the top of scene two. And it's the only alarm inside any of the three generations. Look at this beauty.

[8:25] [panting] Which way? Right. The shutters take out those three moonlight stripes one at a time behind them, which is the corridor plate paying off. He asks which way, she says right. And they cut right with the camera behind them. So her word and the frame agree with each other. And the last generation is one speed ramp into the

[8:43] window smash. Then a first person fall.

[8:56] That slow motion shot has both faces and hundreds of individually spinning shards in the same frame, which is the hardest thing in the entire film. All three generations are done, and the faces match their sheets in every one of them. But one shot inside scene two still comes back wrong. Claude can't watch the video it just made, so every mistake in it is one I have to catch myself. It asks me to play each one back and tell it what went wrong. And that's the only check on the video itself. And the opening close-up of scene two fails. The box in frame doesn't actually match the

[9:22] prop sheet because the proportions are wrong and the lid lining is off. But everything from the shutter slam onwards is fine. So the obvious move is to reroll the whole 15 seconds, but that throws away five approved shots and the shutter slam along with them. Claude regenerates that one shot on its own instead as a standalone insert at 4 seconds. Then I trim it into the 2 and 1/2 second slot in the edit and the extra runtime is there so I can find the cleanest cut point. So I ask it to write the insert prompt and two things change in there while nothing else moves at

[9:48] all. Jewelbox gets named as the single source of truth and the box gets spelled out explicitly rather than left to the tag. Then the exact things that went wrong go in by name as hard exclusions. So the prompt says it must not come back as a ring box or a tall box. The lining is never colored and the necklace is never multistrand. And naming what actually went wrong is the one place

[10:06] where negative wording is worth using. [music] But the framing, the underlight, the whispered line, and the audio locks all carry over unchanged. That's why the cut doesn't show. And as you can see, the box now matches the sheet wide and shallow with the cream silk lid. So scene two keeps its other five shots, and we've paid for a 5-second render instead of another 15. On my own, I'd have just rerolled the whole thing and paid for those 15 seconds twice. That's the whole film approved and still at 1080p. So the upscale is next, and the order you run it in decides what it costs. All four pieces are approved now,

[10:35] which is three scenes and the insert. So, I just tell it all four are approved and to upscale them to 4K. It runs through the MCP inside the same conversation with no download and no re-upload anywhere, and they come back at 4K. Nothing gets enlarged until it's already approved. So, we never pay to upscale a shot that ends up cut. With all four at 4K, the only thing left is the edit notes, and honestly, it's a few plain lines rather than a document. It says to run scene one in full, then drop the insert and trim to about 2 and 1/2 seconds. Then, scene two runs from the

[11:02] shutter slam onwards with its original opening dropped, and then scene three in full. Every single join is a hard cut. And then the music, which is the part we kept out of every generation on purpose. A light comic track runs under scene one on its own. It switches to a chase track the moment the shutter hits the sill, cut dead on that frame with no crossfade. The alarm bell runs from that same frame until the glass smashes and

[11:22] then everything drops out for the fall. And the only thing that crossfades anywhere is the room tone held under both joins so it doesn't jump while the picture cuts hard. Touch nothing. [snorts]

[11:39] It's a nice room.

[11:47] Look at this beauty.

[11:59] Which way? Right.

[12:06] [music]

[12:14] And I didn't have to work any of that order out myself because the prompt wrote those notes as part of the same run and none of that cost anything beyond the credits I already had. The same connector opens every other model on Higsfield and I only used two of them. So open Higsfield through the link below, connect the MCP and run the prompt on your first idea. Thank you for watching and I'll see you in the next

---

About this transcript
Read from YouTube's own caption track and laid out by ViewRank AI (https://viewrankai.com).
ViewRank AI finds the videos already beating a creator's own average on Instagram, TikTok and YouTube Shorts, transcribes them from the audio itself in more than 60 languages, and turns what worked into new ideas and scripts.
Free transcript tools, no account needed: https://viewrankai.com/tools
How to read any video this way: https://viewrankai.com/llms.txt