11:18metricsmule · Tutorial
The Claude Skill That Auto-Writes Seedance 2.5 Prompts
A Claude skill trained on Seedance 2.5's own 35-page prompting guide turns a one-line idea into a shot-by-shot video prompt, no reading required.
August 11thThe argument in one line.
Claude can be turned into a reusable Skill that takes just a song and a few character reference images, then outputs an entire AI music video's treatment, storyboards, prompts, and edit plan in one run.
Read if. Skip if.
READ IF YOU ARE…
- You make AI-generated music videos or character-driven content and need consistent lip-sync, wardrobe, and scene continuity across dozens of short clips.
- You already prompt Claude for an image or video tool by hand, one request at a time, and want to turn that repeat process into a reusable Skill instead.
- You're connecting Claude to an MCP tool like Magnific or Higgsfield and want a concrete look at how the upload and generation handoff actually works.
SKIP IF…
- You want a from-scratch explanation of what a Claude Skill is; this video jumps straight into building one without covering the basics.
- You're looking for live-action video production technique; this is entirely AI-generated character video built from text and image prompts.
The full version, fast.
A creator builds a reusable Claude Skill called 'Music Video' that takes a Suno-made song and a handful of character reference images and outputs an entire AI music video production pipeline in one run: the song cut into aligned 30-second-or-less segments, a treatment describing the character's world and look, an Identity Kit of face and outfit reference sheets, a 4x4 master storyboard plus 3x3 per-segment storyboards, performance and cutaway prompts, and a precise edit plan with beat-synced timestamps. Connected to Magnific's MCP inside the same chat, Claude generates the character sheets, a vehicle reference sheet, and the final clips, then hands back exact in and out points for assembling everything in an editor.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:51
01 · Intro & Hook
Cold open stating the video's promise, accurate lip-sync, character consistency, and scene continuity, followed by a subscribe ask and a tease that future videos will route the same workflow through other AI tools (ChatGPT, Astra, Dots, Muse) and a full multi-agent team.
00:51 – 02:46
02 · Researching the AI Video Model
Before writing any video prompts, the creator asks Claude general context questions about the AI video model's cinematic-realism features, treating the answers as a style foundation rather than content to read directly; Claude auto-saves every generated prompt to an Airtable database, and a Let's AI School community plug follows.
02:46 – 04:05
03 · From One-Off Prompts to a Reusable Agent
Claude is asked for prompts specifically aimed at AI music videos with accurate lip-sync, and the creator pivots from one-off prompting to building a reusable agent, explaining that everything learned so far in the chat becomes the foundation for that system.
04:05 – 06:17
04 · Defining the Character and the Rules
The creator uploads his AI character and a past music video as a style reference, debates with Claude whether to use a storyboard-grid approach or reference images and character locks per segment, and settles it by telling Claude to iterate together rather than demand a final answer immediately.
06:17 – 07:24
05 · Building the "Music Video" Skill
Claude asks four clarifying questions, who sings, is the song finished, what aspect ratio, and how continuity should work, before building a skill called "music video" that handles the entire pipeline; it's built in 43 seconds and tested on an uploaded Suno AI song, which Claude cuts into aligned segments on its own.
07:24 – 08:55
06 · The Production Bible
The skill's full output is explained: a segmented song, a treatment covering world, look, outfits, and locations, Identity Kit prompts with lip-sync mouth shapes, a 4x4 master storyboard plus 3x3 per-segment storyboards, performance and cutaway prompts, and a beat-synced edit plan, all bundled into one production bible.
08:55 – 10:36
07 · Connecting Magnific via MCP
The song is broken into seven segments, and after the creator approves generating everything, Claude explains how to connect Magnific through an MCP connector in Claude's settings, then drives the handoff by opening Magnific's own upload panel for the character reference images since the chat can't upload to Magnific directly.
10:36 – 12:02
08 · Generating the Identity Kit and Car Sheets
Magnific generates the character identity kit along with consistency sheets for the supercar the character drives, locking face, outfit, and vehicle across shots; the creator notes GPT 2.5 Go performed better than the video model specifically for generating character sheets from reference photos.
12:02 – 14:05
09 · Exact Edit Timestamps, Cost, and Assembly
Claude hands back exact in and out timestamps for every B-roll cutaway and the whole video's total generation cost, about 35,550 credits; the finished shots are assembled in Final Cut Pro, and the video closes on Magnific's new experimental built-in video-project editor, which could replace an external editor entirely.
Lines worth screenshotting.
- A single Claude Skill, built once, can take just a song file and a few character reference images and output an entire music video production pipeline: treatment, Identity Kit, storyboards, shot prompts, and edit plan.
- Claude can cut an uploaded song into aligned segments with handles on its own, capping each piece at 30 seconds or less to match most AI video generators' clip limits.
- Asking Claude clarifying questions before building, who sings, is the song final, what aspect ratio, how should continuity work, produced a system tailored to the specific project instead of a generic template.
- The production bible Claude generated included a 4x4 master storyboard, a 3x3 storyboard per segment, performance prompts, 6-10 story cutaway clips, and a precise edit plan with exact in and out timestamps.
- Claude flagged its own uncertainty upfront, estimating lyric timing it couldn't verify and naming exactly which two segments needed to match an existing reference clip, instead of presenting guesses as settled fact.
- A full two-minute AI-generated music video, images included, ran about 35,550 generation credits in total, roughly 2,000 credits per image.
- A content filter blocked one segment's generation twice even though nothing in the prompt was explicit, which quietly added extra cost from the failed attempts.
- GPT 2.5 Go reportedly performed better than the video model for generating character reference sheets from photos, even though the video model handled the actual performance shots.
- Claude can hand off exact cutaway timestamps, down to a tenth of a second, for where to place B-roll like a car or city shot inside the edit, removing manual timing guesswork.
- At least one AI generation platform has started adding a built-in, timeline-based video editor, which could remove the need for a separate tool like Final Cut Pro or CapCut entirely.
One Claude Skill Can Direct a Video
AI VIDEO PIPELINE
Turning a repeat AI-video prompting process into a single Claude Skill lets one upload trigger the whole pipeline, from song segmentation to a beat-synced edit plan, without rewriting instructions each time.
02Researching the AI Video Model
- Before writing any video prompts, the creator asked Claude open general questions about the AI video model's cinematic-realism features, using the answers as a style foundation rather than reading them directly.
- Letting Claude auto-save every prompt it generates to an Airtable database removes the manual copy-paste step from prompt research entirely.
03From One-Off Prompts to a Reusable Agent
- The real pivot wasn't asking for better prompts, it was asking Claude to turn the whole prompting process into a system that could run itself on any future project.
- Everything the reusable system needed came from information already surfaced earlier in the same chat, so the groundwork research wasn't wasted even though it felt unnecessary at the time.
04Defining the Character and the Rules
- Uploading a past music video as a style reference, instead of only describing the look in words, let Claude match an established visual identity without reinventing it.
- Telling Claude to iterate instead of demanding a finished answer immediately turned prompt-writing into a two-way planning conversation before any generation started.
- Before building anything, Claude asked four clarifying questions, who sings, is the song finished, what aspect ratio, and does the world reset per song, which shaped a system fit to this project instead of a generic template.
05Building the "Music Video" Skill
- Claude can cut an uploaded song into its own aligned segments with handles, a capability the creator didn't expect was possible from a chat interface.
- Once the full requirements were answered, Claude built the reusable skill in 43 seconds, after which starting any future chat with the trigger phrase runs the entire system automatically.
06The Production Bible
- Running the skill once produces a single bundled production bible, every prompt, every storyboard, and every timing cue needed for the whole video in one document.
- The skill's output scales with the clip-length limits of the video generator, 30 seconds or less per segment, rather than assuming a fixed video length.
- Claude flagged its own uncertainty upfront, estimating lyric timing it couldn't verify and naming exactly which two segments needed to match an existing reference clip, instead of presenting guesses as settled fact.
07Connecting Magnific via MCP
- Connecting an image or video generation tool to Claude through an MCP connector lets Claude trigger generation directly inside the same chat instead of the creator copy-pasting prompts into a separate app.
- When a platform can't accept uploads directly through the chat, Claude can still drive the handoff by opening the tool's own upload panel and telling the creator exactly what to drop in.
08Generating the Identity Kit and Car Sheets
- A full Identity Kit isn't just a character face sheet, it can also include consistency sheets for key props and vehicles so recurring objects don't shift in appearance between shots.
- Different models performed better at different jobs in the same pipeline, GPT 2.5 Go for generating character reference sheets from photos, a separate video model for the actual performance shots.
09Exact Edit Timestamps, Cost, and Assembly
- Claude can hand back exact in and out timestamps for every B-roll cutaway, down to a tenth of a second, removing manual timing guesswork from the edit.
- A full AI-generated two-minute music video, images included, ran about 35,550 generation credits end to end, roughly 2,000 credits per image.
- Content filters can block a segment's generation for no clear reason, even with nothing explicit in the prompt, which quietly adds extra generation cost from the failed attempts.
- Some AI generation platforms are starting to build a basic timeline editor directly into the tool, which could remove the need for a separate editor like Final Cut Pro or CapCut altogether.
Terms worth knowing.
- Identity Kit
- A set of generated reference images, a character face sheet, outfit cards, and sometimes prop or vehicle sheets, that lock a character's appearance so every generated shot stays visually consistent.
- Production bible
- A single bundled document a Claude Skill outputs that combines every prompt, storyboard, and timing cue needed to generate and edit a full video.
- MCP (Model Context Protocol)
- The connector standard that lets an AI model like Claude call an external tool directly inside a chat, instead of a person copying prompts back and forth manually.
- Claude Skill
- A saved, reusable instruction set Claude follows automatically whenever a chat starts with its trigger phrase, rather than re-explaining the process every time.
- Character lock
- Using a fixed set of reference images plus a written description to keep a character's face and outfit identical across every generated shot.
- Performance prompt
- A generation prompt tagged to a specific audio segment that produces a lip-synced shot of the character singing, as opposed to a cutaway shot with no lip-sync.
Things they pointed at.
01:03toolSeedance 2.5
02:17toolAirtable
06:46toolSuno AI
09:07toolMagnific (MCP)
09:07toolHiggsfield
11:17toolGPT 2.5 Go
12:53toolFinal Cut Pro
Lines you could clip.
00:06
“You're going to be able to create the best music videos you've ever made with accurate lip syncing, character consistency, and scene continuity.”
05:53
“I'll be honest. I didn't know Claude could do that.”
06:17
“Now instantly in 43 seconds, it built the skill.”
07:21
“Each piece is 30 seconds or less. Then it makes a full treatment.”
13:13
“They actually just added something... we have video project editor.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphor
You're going to be able to create the best music videos you've ever made with accurate lip syncing, character consistency, and scene continuity.
And all you're going to have to do, just simply tell Claude what you want. That's it. Today, I'm going to show you how to create a Claude agent that will create, produce, and direct a full music video.
Just like this one right here. So if you're not subscribed, make sure you are.
Because we'll start with this video today, but in upcoming videos, we'll be using all tools like ChatGPT Astra, Dots, Muse, and even deploy a full team agents. that could run everything. Okay, so here we are in Claude, and let me just first show you exactly how this started.
Now, this is simple and maybe even unnecessary, but I think it's important because I first want to provide Claude with some context. So the first thing I did was ask a question sort of irrelevant to what our ultimate goal is. The question is, what is C -Dance 2 .5?
So Claude then provides me with the answer. I'll go through it. I then asked the follow -up question, what features make it stand out from the rest?
Specifically, cinematic realism? ai video be detailed okay so what we're doing is we're just going to get the information it's not even necessary to read this in fact you don't even have to but what you want to do is that we want to leverage the knowledge we're getting from claude because we're then going to use that as a starting block foundation for the style and type of music video that we want because ultimately we do want it to look realistic and have those cinematic tendencies Okay, so then as we scroll down again, it gives us the details.
And then the next thing I want to say is provide me with five prompts that make this stand out. And so you can see it over here again, not really necessary, although it is helpful to get some prompts. So we can go ahead and quickly scroll through all of these.
And you can see in here, speaking of my automations, you can see once I stop right here that you could see where I said add to Airtable. That's something I'll show in a later video where as soon as I said. at Airtable, every prompt that it just provided me here, it's automatically going to put them in one of my Airtable databases.
Want every product, every playbook, and every database, even products not available anywhere else? Then join the Let's AI School community where you get it all. Link in description.
Which now brings us to the next step. Now, here is where we start the real process. What I said was, reflect back to the most impressive and detailed prompts.
Now, I need prompts specifically to generate AI music videos with accurate limp syncing with referenced audio. So now what it's doing is now it's going to give me some example prompts that now you can see how they're created and orchestrated perfectly with multiple shots, you know, all the way to 30 seconds.
And it has those important camera movements, very detailed. So again, we can test one out, but we can quickly scroll through here because the next thing we're going to say is going to be the most important one. for starting the whole entire process and here's what you're going to say now remember you may think i thought we were creating an agent here well we're going to create the agent that's going to do it all for us what we're going to do though first is use a chat like we're doing right now then we're going to leverage all of the information and the references we're putting into this chat tie it up And that is how we're going to create our agent.
It's going to be based on all of the information we're getting, providing, and receiving in this one chat.
So here it is right here and here is where it really started to impress me here is the full prompt now I'll share this prompt via the link in the description, but it's very simple All we're doing is just saying basically what we want. So what I first did is that I added my character right here, Sunkiss.
And then in addition to her, you can see I also uploaded a reference video. Now, this video is another prior music video I made in the past, which in fact, you don't even have to do. Now, what I did was I provided this video just simply as a reference because I wanted my style to sort of match it.
It's always not necessary. You can even add an image. And then in conjunction with this prompt, this is what we're going to get.
And you can see one of the questions I asked was, should we do this with a storyboard grid image for each 30 second video or just continue using references and character locks for each 30 second portion? Now, what's key is the process where I said, let's iterate this together and come up with the best solution. Now, that's one key word I always recommend using.
iterate or iteration because what it's going to eventually end up doing at the same time of creating your music video you're also going to have a brainstorming back and forth process to where ultimately we're going to get the best style and format that we should follow okay and here's what claude said i watched your lv nights cut it already works but here's some things we can basically do.
All right. So here we go. It's going to just tell us what's going to happen.
Let's scroll down. Okay. This is the full iteration process.
So you can see right here, it's going to give us an example, two minute plan, but let's scroll down. And it says questions before I build the templates. This is an important answer.
And it says, who sings? What about the song details, aspect ratio, and then the continuity. Now look at the answers.
Only Sunkissed, which is going to be my character. Number two is what really impressed me. The song, is it finished and available?
If you upload it, I can cut it into sectioned aligned segments with handles myself. I'll be honest. I didn't know Claude could do that.
So I said, number two, I'll let you cut it in the sectioned aligned segments. Okay. And then the continuity is based on my subject and or description.
I decide. Let's do it. Let's build a full system where you do everything you mentioned.
All based on my references, images, character images, and a brief description. Now here is where we build the skill. Build a skill called music video that will handle everything.
It's going to do it. After it does it, now we're going to test it. So here we go.
Now instantly in 43 seconds, it built the skill. I went ahead and saved it. You can see here is the full skill that Claude created.
All right here. The large skill with all of the code and everything like that, it's done. All I have to do in any future chat is just start a chat with right here, music video, and it instantly starts the process, okay?
So let's scroll down. And so here's what we did. Starting this process, I then uploaded a song I created in Suno AI, okay?
The song... was automatically cut up here's what Claude said the song cut into segments so here is exactly how to use this new skill just simply say music video attach a song and then just your image it will instantly do everything the song it's going to be cut up in the segments Each piece is 30 seconds or less.
Then it makes a full treatment. The world, a shared look line used in every prompt, names, outfits tied to segments, all of it. It's going to give us the identity kit.
We're going to get prompts to make character sheets, okay? Then we're going to get a 4x4 master storyboard for the whole video, plus a 3x3 storyboard per segment. Performance prompts, 6 -10 story clips, a full editing plan.
and then everything tied up into one now once we scroll down let's now upload the song we want to use now what i did was i created this song in suno ai and i simply uploaded it with my uploaded audio reference i then just said what i want here it is let's scroll down and let's see exactly what we got this is what really impressed me here it is right here now let's open this up you can see everything we get here is the full thing it's a lot of information But it's so important.
We get all the prompts, all the phases, everything in this one large we can use and save. Now, the other thing we can do with this whole production Bible, we can even use this for future things we want to make, agents, skills, etc. But then it breaks it up.
Here is my song called Starlight Rider. Here's the full song right here. Okay.
It's two minutes and 12 seconds, but now below it, it broke it up into seven segments. So look, segment one, let's go ahead and hit this play button. This is going to be the intro.
Okay. Now that's going to stop at the 18 second mark, which then it brings up segment two. Okay.
This is going to start at that end mark. And now here is the second portion of that song.
Chase the light tonight Okay. So the next thing is I said, let's go ahead and generate the full video.
Okay. From there, it says, do you want me to generate the identity kit now in Higgs field or Magnific or add the prompts to Airtable? So what I would recommend next is to connect this with an MCP.
In this case, I used Magnific MCP. And so if you're unfamiliar with that, make sure you watch this video right here. I'll walk you through the very simple process.
All you have to do is click your settings and then click the connectors in Claude. And then you could see all the ones I have. just simply connect them search for them or even add a custom one with a link very easy to do okay so go ahead and get your MCP set up all that's gonna do is allow you to then use Magnific or Higgs field or whichever one in here and Claude will go ahead and generate everything for you all directly from within this chat once I said right here yes generate identity kit in Magnific let's scroll up here it is here is the full identity kit that it gave us with all of the segments i said yes let's go ahead and do it all now the cool thing about this is that what it's going to ask you to do the workspace can't upload the magnetic directly so it's going to create this drag and drop box right here and that's exactly what i did once i clicked and i dropped in an image of my ai influencer or my character that is how claude was able to create these images of her with magnific so anytime you bring something in here hit continue it will then have access to your character which will then create the full identity kit so now as we scroll down
Now we get her in this look. Look how amazing this actually turned out. It said used Magnific.
Okay. So it says again, upload any more Sunkist references. The three you sent.
plus anything else you can just keep dropping into this box right here. And so if I populate this right here in the Magnific app, you can also see that it generated the car sheets, the grid image sheets of the supercar she's going to be driving in. So we're going to have a few different sheets.
We're going to have a character sheet, a scene sheet, and then we're going to have the grid images of this vehicle she'll be in to guarantee character, scene, and car consistency. And one other thing I want to mention, if you want to use a specific model, you could just say just that.
I said use GPT 2 .5 Go. That way, I want to make sure you generate the character sheets with GPT 2 .5. It's usually going to perform a little bit better when you're using reference images to see Dance 2 .5 videos.
And now here it is. Look. We get more images right here.
Now as I scroll down, it's giving us the full clips and the full editing process. So here's the first clip. Everything's broken down.
Okay, look at this. Let's scroll down. I said, give me the timestamps of where I should add the B -roll of the car, city, etc.
Now if I scroll back up, it says all eight story cutaways are done and saved in your Sunkist folder. Here's where each one goes in the edit. So yes, you do have to open up some form of editing tool.
It could be anything. Just something to put. the clips side by side, but it tells you specifically in each timestamp very precisely, like 0, 3 .8, where to put the actual overlaps or where to put any of the transitions.
Extremely helpful, okay? Then I said, give me the timestamps of where I should add the specific B -roll, the car, and it gives it to me perfectly right here, the length, timeline, in and out of each of the B -roll clips. Now, let me just show you very quickly how this all looks.
in Magnific directly, here is everything right here, okay? Everything that it created. We could see all of these images were done with the MCP and Claude, the scenes, the cars, and then we get the individual videos all created from her driving it to her close -up shot singing to the actual car in the garage.
It's all right here. And now the next thing, once you piece this together, like I just simply used Final Cut Pro, but you can use like anything that's easier. But the other thing you may not know about, if I come back over here into Magnific, watch this.
They actually just added something. Let's go to all tools. And now if I select specifically video right here, click this, look at this.
We have video project editor. It's experimental, but now if I go to a new video project, okay, open this up, look what we get. All of our items are landed here.
We get a full timeline. We can actually start the full editing process here and we won't have to use an outside tool like a Final Cut Pro or CapCut. But nonetheless, it's extremely easy to put the clips side by side.
And if you like this video, make sure you also check this video out right here. It's in the description. I'll see you next time.
The Hook
The bait, then the rug-pull.
Most AI music videos still break on the thing a viewer notices first, a face that drifts between shots or lips that don't quite track the words. The creator's fix wasn't a sharper prompt, it was building Claude a reusable Skill that remembers an entire production end to end, from a raw Suno file to a finished edit plan with exact in and out points.
Named ideas worth stealing.
07:50list
The Music Video Skill's output
- Song cut into 30-second-or-less segments
- Treatment: world, shared look, outfits, locations
- Identity Kit: character face sheet + outfit cards
- 4x4 master storyboard + 3x3 per-segment storyboard
- Performance prompts (lip-sync) + story cutaway clips
- Edit plan: cutaway over every segment join, cuts on the beat, one color grade
The full set of deliverables the custom Claude Skill produces from a single song-plus-reference-images upload, bundled into one production bible.
Steal forany repeat AI-video workflow worth turning into a one-shot Claude Skill, like a VOD sizzle or reel pipeline.
05:43list
Questions before building the templates
- Who sings, and does the format need single lip-sync or a duet?
- Is the song finished and available to cut into segments?
- What aspect ratio, 16:9 or 9:16?
- Does the look and world carry over between songs or reset each time?
The four clarifying questions Claude asked before building the reusable skill, so the system matched this specific project instead of running a generic template.
Steal forany Claude Skill that should branch its behavior per project instead of assuming one fixed setup.
How they asked for the click.
VERBAL ASK
02:40product
“Want every product, every playbook, and every database, even products not available anywhere else? Then join the Let's AI School community where you get it all. Link in description.”
Soft mid-video plug dropped right after a free-value beat (prompts auto-saving to Airtable), framed as a value-add rather than a hard sell.
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
AFFILIATECommission earned if you click.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
