How to Create High Quality AI Videos With Claude — Transcript
Full transcript
- 0:00Making AI videos just got a whole lot
- 0:01simpler because you can now run the
- 0:03entire thing from a single Claude chat.
- 0:05You describe what you want and Claude
- 0:07generates it in the background using
- 0:08Higgs Field.
- 0:28I used it to build this full cinematic
- 0:30underwater film and I'm going to walk
- 0:32you through the whole thing. If you've
- 0:33ever made AI videos, you know the
- 0:35struggle. You write prompts in one tab,
- 0:37generate in another, download,
- 0:39re-upload, and half your session is just
- 0:41moving files between tools. Higgs Field
- 0:43fixed that with something called MCP,
- 0:45which basically lets Claude plug
- 0:46straight into Higgs Field so Claude can
- 0:48generate images and videos itself right
- 0:50in the chat. MCP stands for Model
- 0:52Context Protocol and in plain terms,
- 0:54it's a standard way for AI tools to talk
- 0:55to each other. So once the connection is
- 0:57established, Claude can see Higgs
- 0:58Field's tools and run them as if they
- 1:00were its own. And the setup takes about
- 1:022 minutes, so let's get it done. If you
- 1:03want to follow along, the Higgs Field
- 1:05link is in the description. The page
- 1:06shows you the setup steps along with the
- 1:08exact link to copy. [music] Now over in
- 1:10Claude, go to settings, then connectors,
- 1:12and click the little plus icon to add a
- 1:14custom connector. Name it Higgs Field,
- 1:16paste that link in, and hit add. Once
- 1:18it's connected, a permissions panel pops
- 1:20up where you pick which tools Claude is
- 1:21allowed to use. And at minimum, you want
- 1:23generate image and generate video turned
- 1:25on. And the full list goes way past
- 1:27those two. [music] There's background
- 1:28removal, reframing, presets, character
- 1:30tools, nearly everything the platform
- 1:32can do, all callable from the chat.
- 1:34[music] I'd keep that list tight, too,
- 1:35because every tool you enable is
- 1:37something Claude can run on its own. So
- 1:39only switch on what you actually plan to
- 1:40use. Now let's test it and I'll paste in
- 1:43this prompt. And when it runs, you
- 1:44actually see the Higgs Field job appear
- 1:46right inside the chat. It processes
- 1:48there and the finished image drops into
- 1:49the conversation like any other reply.
- 1:51We get a golden retriever with full
- 1:53eagle wings flying above the clouds with
- 1:54that cinematic blue tint I asked for in
- 1:564K. And notice I never opened Higgs
- 1:58Field because the prompt itself
- 2:00specifies which model to use. And if you
- 2:02break it down, that one sentence gave
- 2:03six separate instructions. The subject,
- 2:05the color grade, the shot style, the
- 2:07model, the quality, [music] and the
- 2:08aspect ratio. I made the test 4x3 on
- 2:10purpose, too, because ratio control is
- 2:12the easiest thing to check at a glance.
- 2:14And the film we're making next is 16x9,
- 2:16so switching between them is nothing
- 2:18more than changing the words in the
- 2:19sentence. [music] And that's the core of
- 2:20this whole workflow. Your settings go
- 2:22inside the sentence, so instead of
- 2:23clicking through menus, you just say
- 2:24what you want, the model, the quality,
- 2:26the resolution, the ratio, and Claude
- 2:28handles the rest. But a test image is
- 2:30the small [music] stuff, because the
- 2:31real reason we connected these two is to
- 2:33make an actual film. By the way, if you
- 2:35haven't opened Higgs Field yet, I'd
- 2:36really recommend opening it from the
- 2:38link in the description and following
- 2:39these steps along with me, because it
- 2:41makes the whole thing much easier to do.
- 2:43And you'll walk away from this video
- 2:44with a finished film of your own. We're
- 2:45making a 36-second underwater short film
- 2:48in first-person view with three videos
- 2:50of 12 seconds each. And the story runs
- 2:52from calm exploration to a chase to a
- 2:53discovery at the end. And first person
- 2:55is a deliberate choice because the
- 2:57diver's hands are in every single shot,
- 2:59and that's exactly what makes them easy
- 3:00to keep consistent. And the whole thing,
- 3:02from the concept to the final footage,
- 3:04never leaves the conversation. And
- 3:05instead of prompting each piece one at a
- 3:07time, I'm giving Claude the entire
- 3:08production plan in one message, so I'll
- 3:10paste this in. It's a long one, but the
- 3:12structure is simple. It splits the job
- 3:13into three stages: the reference frame,
- 3:15the sequence breakdown, and the
- 3:17generation. And it tells Claude to stop
- 3:19and wait for my approval after every
- 3:20stage. Those approval stops matter
- 3:22because without them, Claude would run
- 3:24all three generations in one go. And if
- 3:26the first video is off, everything after
- 3:27it is off, too. And you can reuse this
- 3:29exact structure for any film you want to
- 3:31make. Keep the three stages and the
- 3:32approval stops, and just swap out the
- 3:34story in the middle. Stage one is the
- 3:36reference frame, and this is the film's
- 3:37consistency trick. Before any video gets
- 3:39made, Claude generates one image of the
- 3:41diver's hands and gloves underwater, and
- 3:43that exact image becomes the anchor for
- 3:45all three videos. So the gloves, the
- 3:47equipment, and the water tone stay
- 3:48identical from start to to And the
- 3:50prompt specifically says this is not a
- 3:52scene, just the hands and gloves against
- 3:54open water." Because an anchor works
- 3:55best when there's nothing in it except
- 3:57the things that need to stay consistent.
- 3:59The first version comes back honestly
- 4:0090% there. The gloves, the water tone,
- 4:03and the light shafts all look right. But
- 4:05there's a dive watch on the left wrist
- 4:06that I don't want. So I'll tell it to
- 4:08remove the watch and put something else
- 4:10on his left hand. And the second version
- 4:11comes back with the torch in his left
- 4:13hand and the beam cutting through the
- 4:14water. This frame is exactly what an
- 4:16anchor should be. The charcoal gloves,
- 4:18the torch with its beam visible, deep
- 4:20teal water with light shafts coming down
- 4:22from above. Simple enough to match in
- 4:23every shot and specific enough that you
- 4:25notice if it drifted. So I'll approve it
- 4:27and tell Claude to move to stage two.
- 4:29Now Claude writes the full prompt
- 4:30[music] for each video and shows me all
- 4:32three together before generating
- 4:33anything. So I get to approve the shape
- 4:35of the whole film at once instead of
- 4:37finding problems one generation at a
- 4:39time. And these prompts are worth a
- 4:40closer look because every single one
- 4:42starts with the exact same description
- 4:44of the hands, the charcoal gloves, the
- 4:45wet suit sleeves, the torch with the
- 4:47yellow grip ring. And that repeated
- 4:48block is what keeps the diver identical
- 4:50across all three videos. Then each video
- 4:52is broken into five numbered shots with
- 4:54a time on each one. Two seconds here,
- 4:56three seconds there, so every 12-second
- 4:57video is really a five-shot sequence
- 4:59with its own edit already built in. Take
- 5:01the first one. Shot one pushes through a
- 5:03rock arch, shot two tilts up into a
- 5:05spiraling school of silverfish, shot
- 5:07three grazes a sea fan coral, shot four
- 5:09glides through a field of jellyfish, and
- 5:11shot five brings in the sea turtle and
- 5:12that dark shape at the end. All time
- 5:14down to the second. And one more detail
- 5:16in the master prompt is doing a lot of
- 5:17work here. Each video has to end on a
- 5:19frame that cuts into the next one and
- 5:20that's [music] what makes three separate
- 5:22generations feel like one continuous
- 5:23film instead of three pieces stuck
- 5:25together. And Claude closes every one of
- 5:27them with the same color grade line,
- 5:28deep teal blue, cinematic, which is the
- 5:30third consistency layer after the hands
- 5:32and the settings. [music] All three run
- 5:34on SeaDance 2.0, 12 seconds each, 720p,
- 5:3716 by 9 with ambient sound and Foley
- 5:39only, no music. So the only sound in the
- 5:41raw footage is the regulator breathing,
- 5:42the bubbles, [music] and the water
- 5:43itself because the music gets added in
- 5:45the edit where we control it scene by
- 5:47scene. And the 720p is on purpose, too,
- 5:49because the last step of this workflow
- 5:51upscales everything to 4K. So, there's
- 5:53no reason to spend extra credits on
- 5:55resolution now. The prompts all look
- 5:56good, so I'll approve them, and Claude
- 5:58starts generating. Video one first, then
- 6:00it shows me the result and waits, then
- 6:01video two, then video three. That pause
- 6:03between generations runs on the same
- 6:05logic as the approval stops. Every video
- 6:07costs credits, so you check the one that
- 6:08just finished before the next one spends
- 6:10anything. And the three come back
- 6:11exactly on story. Calm exploration with
- 6:13a sea turtle and a dark shape drifting
- 6:15past at the end. Then the chase through
- 6:17the coral tunnel, then the chest on the
- 6:18seafloor that does not stay shut. The
- 6:20gloves, the torch, and the water tone
- 6:22stay the same through all three, which
- 6:23is exactly what that reference frame was
- 6:25for. And the quality is honestly right
- 6:27up there with generating on the platform
- 6:29directly, which makes sense, because
- 6:30Claude isn't generating anything itself.
- 6:32It's running the same C-DANs 2.0 you'd
- 6:34use in Higgs Field. Now, there's one
- 6:36part of this that Claude can't do for
- 6:37you, and that's the final edit, because
- 6:39stitching the clips into a single film
- 6:41is still a manual job. So, download the
- 6:43three videos, drop them [music] into
- 6:44CapCut, editor you use, and line them up
- 6:47in story order. Then trim the edges of
- 6:49each video, because AI footage always
- 6:51has a rough second or two around the
- 6:52cuts, and trimming those edges is what
- 6:54makes the transitions feel smooth. After
- 6:56that, layer your music to match each
- 6:58part. Something calm for the
- 6:59exploration, something driving for the
- 7:01chase, something uneasy for the
- 7:02discovery. And one thing to know here,
- 7:05ending up shorter than you generated is
- 7:06completely normal. A 36-second
- 7:08generation usually trims down to around
- 7:1030, so always generate more footage than
- 7:12your target, because that extra is your
- 7:14room to edit. And export the final cut
- 7:16as a 720p video. But a 720p file is not
- 7:19what you want to publish, and fixing
- 7:20that is the one step people skip. C-DANs
- 7:22generates at 720p, which looks fine in
- 7:24the edit, but soft on a big screen. So,
- 7:26the last move is taking the whole film
- 7:27up to 4K. And this happens right back
- 7:29inside Claude, so I'll upload the
- 7:31finished MP4 into the same chat and
- 7:33paste this in. The settings go inside
- 7:34the sentence again. The source is our
- 7:36720 export, the provider is Topaz Video
- 7:38AI, the model is Starlight Precise 2.5,
- 7:41the target is 4K, and the aspect ratio
- 7:43stays on auto, so nothing gets
- 7:44stretched. And Claude does the entire
- 7:46thing on its own. It uploads the file to
- 7:47Higgsfield, checks that the upload went
- 7:49through, confirms the job, and runs the
- 7:51upscale through Topaz Video AI with the
- 7:53Starlight Precise 2.5 model, taking our
- 7:55720p film to full 4K with the aspect
- 7:57ratio preserved.
- 8:28And the difference is obvious. The water
- 8:30texture, the particles, the torch beam,
- 8:32everything holds up on a big screen now.
- 8:34And that order, generate at 720p, edit,
- 8:36then upscale once at the end, is the
- 8:38cheap way to do this, because editing
- 8:39low-quality files is faster on any
- 8:41machine, and you only pay for 4K a
- 8:43single time on the finished cut. So, the
- 8:45whole film took one Claude conversation
- 8:46and one editing session. Claude
- 8:48generated the anchor image, wrote the
- 8:50prompts, made all three videos, and
- 8:51finished it in 4K, and the only thing I
- 8:53did by hand was the cut. And the same
- 8:55setup is how I run my image jobs, video
- 8:57jobs, and upscales now. One chat, plain
- 8:59English, approve the steps, collect the
- 9:01result. And the part that sold me on
- 9:03this setup is how little I ever left the
- 9:04chat, because from the first test image
- 9:06to the 4K export, Claude ran all of it
- 9:08through one connection. So, when you're
- 9:10ready to build your own, the Higgsfield
- 9:12link is in the description below. Thank
- 9:13you for watching, and I'll see you in
- 9:15the next one.
About this transcript
This page contains the full transcript of How to Create High Quality AI Videos With Claude by Parker Prompts, generated from the public captions YouTube serves with the video. The transcript has 2,011 words across 290 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.