A screen-recorded walkthrough of building a custom Claude Code skill that watches any viral video, breaks it into timestamped beats, and recreates it with Higgsfield's Seedance 2.0.
A single custom Claude Code skill can turn any viral video into a repeatable remix pipeline by pairing Fable 5 as a planning advisor with Sonnet 5 as the cheaper executor, cutting token cost by roughly 80 percent.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You use Claude Code and want a concrete example of advisor mode (Fable plans, Sonnet executes) to cut API costs on agentic tasks.
You're building or maintaining custom Claude Code skills and want to see a real plan-mode review-and-revise cycle.
You make short-form AI video content and want a repeatable process for reverse-engineering a viral video into a new prompt set.
You're evaluating Higgsfield's MCP/CLI for AI video generation and want to see it driven end-to-end from a terminal agent.
SKIP IF…
You don't use Claude Code or have no interest in agentic terminal workflows — this is a builder-facing tutorial, not a finished product demo.
You're looking for a no-code tool; this requires installing a CLI, an MCP connector, and reviewing a generated skill.
TL;DR
The full version, fast.
Claude Fable 5 is powerful but expensive at $10 per million tokens, so the creator uses it only as an advisor (the planning brain) while Sonnet 5 executes the actual work, cutting cost by about 80 percent. He installs a free community skill that lets Claude watch any video (extract frames, transcribe audio) and gives Fable a goal: build a new skill that watches a viral video, breaks it into camera-move and character-action beats, and writes a Higgsfield Seedance 2.0 video prompt in 15-second timestamped chunks. Claude drafts a plan in plan mode, the creator reviews and edits it (correcting resolution and aspect ratio), then approves auto mode. Claude builds the working skill, tests it against a fresh terminal session, and when Seedance fails, Claude automatically falls back to Cling 3.0 to still produce a video. The core takeaway: give Claude a clear goal and success state, review the plan before execution, and let it self-correct when a tool fails.
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.
Sets up the premise: his community asked for a viral AI video generator, built by recreating an existing 1.8M-view video.
00:43 – 01:00
02 · How to Access Fable 5
Notes Fable 5 is the most capable model because it excels at figuring out what you want from a stated goal.
01:00 – 01:21
03 · How Fable 5 Works Best
Explains the give-it-a-goal-and-walk-away philosophy, illustrated with an on-screen diagram, and flags Fable's cost.
01:21 – 01:43
04 · How to Cut Token Costs
Fable costs $10/million tokens — 2x Fable 4.8, 5x Sonnet 5 — so the plan is to use Fable as advisor and Sonnet to execute, saving about 80%.
01:43 – 02:02
05 · How to Enable Advisor Mode
Advisor mode is terminal-only for now; shows Sonnet 5 already selected as the active model via the /model command.
02:02 – 02:34
06 · How to Install Video Skill
Installs a free community GitHub skill (by Brad Automates) that lets Claude extract frames, transcribe, and understand any video by pasting the repo link.
02:34 – 03:52
07 · How to Watch Any Video
Demonstrates /watch on a Goku vs Thanos AI fight video, returning a timestamped beat breakdown of actions and camera moves.
03:52 – 04:09
08 · How to Set Fable Advisor
Clears the session and runs /adviser to switch Fable 5 into the advisor role while Sonnet executes.
04:09 – 05:04
09 · How to Define Skill Goals
States the goal to Fable: build a global skill that watches a video, breaks it into beats with camera/action/style notes, and generates 15-second Higgsfield Seedance 2.0 prompts with timestamps.
05:04 – 06:56
10 · How to Install Higgsfield MCP
Walks through installing Higgsfield both via CLI (terminal/IDE) and MCP connector (desktop app): copy connector link, add custom connector, sign in.
06:56 – 07:48
11 · How to Use Higgsfield Video
Tours the Higgsfield video UI inside Claude (Seedance 2.0): prompt field, duration, aspect ratio (16:9 vs 9:16), and resolution controls that affect credit cost.
07:48 – 08:23
12 · How to Review the Plan
Fable returns the full plan mode game plan for the new global skill, matching the stated goal.
08:23 – 08:43
13 · How to Adjust Video Settings
Catches that the plan defaults to 1080p/16:9 and edits it in plan mode to require 720p and 9:16 for cost savings before approving.
08:43 – 09:58
14 · How to Auto-Build the Skill
Approves auto mode; Claude builds the skill end-to-end, and the creator questions two ambiguous steps in the plan, prompting Claude to clarify and tighten them.
09:58 – 10:58
15 · How to Test the Skill
Tests the finished /viral-remix skill from a fresh terminal; Seedance fails on every attempt, so Claude auto-falls-back to Cling 3.0 and still produces a usable remixed video.
Atomic Insights
Lines worth screenshotting.
Claude Fable 5 costs $10 per million tokens — twice Fable 4.8 and five times Sonnet 5 — so pairing it as an advisor with Sonnet as executor can cut cost by about 80 percent.
Advisor mode, where Fable plans and a cheaper model executes, is currently only accessible from the terminal, not the desktop app.
A single free GitHub skill can give Claude the ability to watch any video by extracting frames, transcribing audio, and handing both to the model.
Installing a Claude Code skill can be as simple as pasting its repo link into Claude and asking it to install itself.
Giving Claude a goal with an explicit success state, rather than step-by-step instructions, is what advisor mode is designed to work from.
Plan mode lets you review Claude's proposed approach and edit specific parameters, like forcing 720p and 9:16 instead of 1080p and 16:9, before any execution happens.
Dropping video resolution from 1080p to 720p in Higgsfield cuts generation credit cost by 50 percent.
Seedance 2.0 clips are capped at 15 seconds, so longer source videos get broken into multiple 15-second chunks and stitched back together.
When Claude isn't sure whether an approach will work, asking it directly and getting it to explain its reasoning is more productive than guessing.
When the primary AI video model (Seedance) failed on every attempt, Claude automatically substituted a different model (Cling 3.0) to complete the task rather than stalling.
Cling 3.0 has a 10-second clip cap versus Seedance's 15 seconds, so a mid-pipeline model swap can quietly change output length.
A reusable Claude Code skill turns a one-off manual process (watch video, write prompt, generate clip) into a slash-command workflow that runs in a couple of minutes on any new viral video.
Takeaway
Split reasoning from execution to cut agentic AI costs.
WHAT TO LEARN
Routing planning to a premium model and execution to a cheaper one, and reviewing the plan before approving it, is what makes expensive agentic AI workflows both affordable and reliable.
01How to Plan Viral Videos
Building from what an audience explicitly asks for, rather than guessing at demand, is a low-risk way to pick what to build next.
03How Fable 5 Works Best
Stating a clear goal and success condition, rather than a sequence of instructions, is what lets an agentic model plan and act independently.
04How to Cut Token Costs
Cached or not, splitting the expensive reasoning model from a cheaper execution model on the same task can cut total token cost by roughly 80 percent.
07How to Watch Any Video
Extracting frames plus a transcript is enough for an agent to reconstruct a video's beat-by-beat structure without watching it 'live'.
09How to Define Skill Goals
A goal is easier for an agent to act on when it specifies the exact deliverable format (timestamped 15-second prompts) rather than just the end result.
11How to Use Higgsfield Video
Lower resolution settings can cut generation costs by half with an acceptable quality tradeoff for short-form remix content.
13How to Adjust Video Settings
Reviewing a generated plan before execution catches concrete, fixable mistakes — like wrong resolution or aspect ratio defaults — before they cost money to run.
14How to Auto-Build the Skill
When you're unsure whether an agent's plan will actually work, asking it to explain its reasoning is more useful than guessing or re-prompting blind.
15How to Test the Skill
A capable agent can recover from a failed tool or model automatically by substituting an equivalent alternative, so a single failure doesn't have to stall the whole pipeline.
Packaging a manual, repeatable process into an installable skill turns a one-off task into a reusable command for every future instance of the same job.
Glossary
Terms worth knowing.
Advisor mode
A Claude Code terminal feature where one model (e.g. Fable 5) plans and reasons while a separate, cheaper model (e.g. Sonnet 5) executes the actual tool calls, reducing token cost.
Plan mode
A Claude Code mode (toggled with Shift+Tab) where Claude proposes a step-by-step plan for a task and waits for human review and edits before executing anything.
Higgsfield
A platform providing access to AI video and image generation models, usable via a web app, an MCP connector for Claude's desktop app, or a CLI for terminal/IDE use.
Seedance 2.0
A Higgsfield video generation model capable of 15-second clips, used here to recreate viral videos scene-by-scene.
Cling 3.0
An alternate Higgsfield video generation model with a 10-second clip cap, used as an automatic fallback when Seedance fails.
Claude Code skill
A reusable, installable capability for Claude's terminal agent (Claude Code) that packages a specific workflow, such as watching a video or generating video prompts, behind a slash command.
MCP
Model Context Protocol — the connector standard that lets Claude's desktop app call external tools and services, such as Higgsfield, directly without opening a browser.
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.
17px
metaphor
Claude Fable five has completely changed the way people are building with AI. In one of my last videos, I searched the Internet to bring you 28 of the best use cases for Fable five, but just because you can build something doesn't mean that you should. So for this video, I went into my community of over 6,700 members and I asked them, if I could build one thing for you, what would it be?
And they almost all said the same thing, a viral AI video generator. So in this video, I'm going to show you how to use Quad Fable five to create viral AI videos the simplest way that I know how, by recreating what's already working. This video has 1,800,000 views and I wanna see if we can use Quad to build a custom skill to recreate any viral video we find on the internet.
It'll watch the video, create the prompts for us, and generate the whole thing in just a couple of minutes. So focus in, close all your open tabs, and let's build. If you wanna get access to the skill we built today already done for you, just check the link in the description.
Fable five is the most powerful model on earth right now, and it's really because of this one concept. It is incredibly good at figuring out what you want it to do. All you have to do is give it a goal, like what do you want it to achieve, and then basically walk away and it's done.
But there is one big problem with it. Fable five is expensive. At $10 for every million tokens, it's twice as expensive as 4.8, and it's five times more expensive than Anthropic's other latest model, Sonic five.
So today, we're going to use a little hack to save you on token cost. We're actually going to use Fable as the adviser, kind of the brains behind the entire operation, but we're going to use Sonic to execute on any of the tasks that we give it. This is going to save you, like, 80% of your cost.
I'm gonna be doing this inside of the terminal today because you can only access advisor mode from the terminal currently, although I imagine it's gonna be rolling out to the desktop app soon. If you wanted to use Fable inside of the desktop app to do everything, you can just come in here and select Fable five to get started.
So I'm gonna fire up Claude, and currently you can see I've already set the model to Sonnet five. If you wanna check, you can just come in here, type in model, and select the model. Now out of the box, Claude can't actually read videos very well, but there is a free GitHub repo that gives Claude the ability to watch any video.
It extracts the frame, it transcribes it, and it hands it all to Claude. This is the Claude video skill by this guy, Brad Automates, and honestly, it is awesome. It's super easy to install these type of skills with Claude.
Literally, just copy the link to it, come over to Claude, paste it in, and say, hey, Claude, install this for me. And in just a couple seconds, you'll be set up all ready to go to watch any video on the Internet. I've already installed this, and to watch any video now, all you have to do is use the slash watch command and then paste in the URL.
So I'm just going to paste in a link to our Thanos versus Goku video and see what we come back with just to show you how this works before we start building our skill. The whole reason that we build a skill is so that we have a really consistent repeatable process so that we know anytime we come across a viral video that we wanna remix for ourselves, we can do it in just a couple seconds.
So the way that I'm thinking that our skill is going to work is going to be really simple. All we have to do is give it a URL. It's going to extract the frames from the video.
It's then going to generate a brand new video prompt for a new video for us, and then we're going to create our video using Higgs field. So once we get this set up, the whole idea is that all you have to do is give the system a URL, and you come out with a brand new video, all set to go viral. So check this out.
This is an AI generated fight between Goku and Thanos. Here's what happens. Goku's in Thanos brawl, grappling hand to hand.
It even gives you the time stamps, then Thanos is going to crouch to attack between one and four seconds, then he grits his teeth, etcetera, etcetera. So we can see that this gives us the entire outline of the video. So I did this just to show you how this works.
I'm going to clear out everything right now, and what I'm gonna do is type in slash adviser to set the adviser to Fable five. And now we can actually have Fable act as the brains behind everything while Sana executes the task much cheaper for us.
And so now we can just tell Fable the type of skill that we wanna build. And what I'm gonna do is I'm going to give Fable a goal because Fable works best when you give it a success state. So what is the ultimate goal you are trying to achieve?
I want to build a new global skill that is designed to create viral AI videos. First, it'll use the slash watch skill to watch the video at the URL that I gave it. It'll break it down into really specific beats, noting the character action in any of the camera movements and the style of the video.
And then I need you to generate a video prompt for Higgs Field using CDance two point o. These should be fifteen second clips broken out into different sections with time stamps so that by the end, we are going to be remixing the viral video that we gave you. So I'm gonna send this off, and let's see what Fable comes back with.
I stopped it right here because one thing I wanna do is I actually wanna hit shift tab on the computer a couple of times to see that we have plan mode on. So Fable is actually going to create a game plan for us first so that we can review it as the human before it starts building. And I'll just say, please continue.
So these systems are super smart, and we can see Fable actually spun up a couple of sub agents for us. The first agent is exploring the watch skill that we already installed, and the second agent is exploring how to create new skills. While that's cooking, let me show you how to install Higgs Field in the terminal and in the desktop app.
If you're unfamiliar, Higgs Field is a platform that gives you access to all the latest video and image models and really allows you to do some incredible things. What's cool is it's also super easy to install. Just come up to the top here and go to MCP and CLI.
You don't really need to know what those means. I will let you know it's recommended that you install the MCP in the desktop app, but it is recommended that you install the CLI if you're using terminal or an IDE like Versus Code. To install this in the terminal, literally just copy this command, open up a brand new terminal window, and paste this in.
It's going to ask you to sign in with your Higgs field name and password, and you're good to go. For the MCP inside of the desktop app, you just want to copy this link to the MCP, come over here into the app, click the plus, go into connectors, manage connectors, and click the plus up here, add custom connector, name this Higgs Field, paste in the link, click add, sign in with your name and password, and now you can use Higgs Field directly inside of the desktop app.
I already have this installed here, and you can see this has access to generating all of the images and video. And what's great about this is now Quad does all the hard work figuring out how to best prompt the system and can just call Higgs Field directly without you ever opening a browser. So this is actually cool.
Quad is asking me, do I want to recast the video with a different subject or brand so, like, I could put myself in the same action, or do I wanna create a direct visual recreation? If If you're trying to create viral video content, like remixing what's already working, like, generally tends to work best. So right now, I'm just gonna say direct visual recreation, but you can always adjust this later as you like.
One of the things that I love about Quad and one of the questions that people ask me all the time is like, should I do this or can I try that? And I always tell people, like, yes. Like, you can't really break the system.
If there's something that you wanna try or a variation you wanna explore, just, like, tell Quad what you are trying to achieve, and it's gonna figure it out for you. If you've never worked with Higgs Field inside of Cloud, I just wanna show you some of the things that you have access to so you can think about it the right way.
I just went up to video in the top and came into Cdance two point o. And, basically, there's just a couple of things that you need to understand. Like, this is the area where you give it a prompt.
You can change the duration of the video. You can change the aspect ratio. So 16 by nine is widescreen.
If you wanna make this vertical for shorts, nine by 16. You can even change the resolution, which is actually going to save you a lot on credit.
So if you come down from ten eighty p to seven twenty p, you're already saving 50% on cost. So these are the same parameters that you have access to through club. So this is the full game plan that Fable five just created for us.
So I want a new global skill that takes a viral video URL, watches it, breaks it into precise time stamped beats with the character action, the camera movement, the style, and generates a Higgs Field CDance two point o video prompt as fifteen second time stamped clips that recreate the source video. And this is the full pipeline on how it's going to execute the task.
There's one thing I do wanna note here that goes back to what I was just showing you in Higgs field. You can see here that right now it says the resolution is ten eighty p, and I just wanna make sure that we're only using seven twenty p to save on cost and that we have the vertical nine by 16 aspect ratio. And so this is why you do plan mode first so that you can review it before Claude just starts working on stuff.
So I'm just gonna tell Claude what I want to change down here, and I'm just gonna say make sure we're using seven twenty p resolution and nine by 16 aspect ratio. So we can see that Quad changed the plan here to nine by 16 aspect ratio and a resolution of seven twenty p.
This looks good to me. I'm gonna say yes. Use auto mode.
Do your thing. So this finish, and it really didn't take all that long at all. It says all tooling is live.
It built the viral remix skill. It watches the video. It extracts the precise beats.
It writes a time coded prompt for us. It renders each window with the original frames and concat, and then it stitches all the clips together if you have them. I do have a couple of questions about this.
So if you're unsure about something, just ask Claude your questions. I have a question on number four. What do you mean by render each window with the original frames?
And then for number five, my understanding was that we would be creating, like, a fifteen second clip inside of Cdance with the seconds time stamped out. And so that way, really, Cdance is going to be doing all the stitching together for us, and so I'm not sure why step five exists. Cool.
So this answered my questions for me. And, basically, this is going to work for a short video like fifteen seconds or a longer video. So if it's a longer video longer than fifteen seconds, Claudius is going to break it out into fifteen second chunks, which is great because the limit for C Dance is fifteen seconds.
C Dance is going to stitch everything together unless the clip is longer, in which case it'll create multiple sequences in C Dance for us and then stitch all of those together to create a longer format video. And it even acknowledges that the distinction wasn't clear in the skill, and it asked me if I wanna tighten the steps, And I'm just gonna say tighten them because we might as well have this thing as tight as we can get it.
So we should be good to go to test this out. So I'm just going to open up a brand new quad terminal just so that we could test this totally from scratch, and I believe our skill was viral remix. So I'm just gonna type that in.
Come back into our link over here, paste this in, and wish me luck. This finished, but one weird thing did happen that I just feel like I should tell you about. So CDance failed on every attempt even when I was using really generic prompts.
It's just a problem with the model, and this happens with the AI image and video models from time to time. They usually get back up and running within twenty four hours. So Quad is super smart, and so what it did instead is it switched from Cdance to Cling, and so the video we actually ended up generating was using Cling three point o instead Cdance.
Cling has a ten second cap, but the results are still pretty good. So let's check this out, and next time Cdance is back up and running, we can just try it again.
So I'd say this is a pretty good start. If you wanna get access to the skill already built for you and join thousands of others learning how to build cool things with quad code, just check the link in the description. If you wanna see how to take our AI video generator one step further and remix the formats from an entire viral channel, check out part two right here.
I'll see you over there.
The Hook
The bait, then the rug-pull.
The creator asked his 6,700-member community what he should build with Claude Fable 5, and the answer was unanimous: a viral AI video generator. What follows is a screen-recorded build of a custom Claude Code skill that watches any viral video and hands back a ready-to-generate remix.
Frameworks
Named ideas worth stealing.
01:36concept
Advisor + executor cost split
Fable 5 plans (advisor)
Sonnet 5 executes (cheaper tool calls)
Route planning/reasoning to the expensive, most capable model and route execution to a cheaper model to cut total token spend by roughly 80%.
Steal forany agentic Claude Code workflow where the task involves both heavy reasoning and repetitive tool execution
CTA Breakdown
How they asked for the click.
VERBAL ASK
10:39product
“If you wanna get access to the skill already built for you and join thousands of others learning how to build cool things with quad code, just check the link in the description.”
Soft CTA delivered at sign-off, pointing to a $9 Skool community product bundling the finished skill; also teases a part-two video.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
Every night at 1am, 3am, 4am, and 5am, a set of Claude Code agents build something, grade themselves against a rubric, and write their own rules before the next run.
Andrej Karpathy says the best way to brief an AI coding agent is a messy ten-minute voice ramble — so a creator tests it live by building a naturopath website from a single, uninformed viewer request.
A former art director breaks a single sentence into a seven-piece "goal prompt" and gets Fable 5 to autonomously build a playable game, a 3D website, and a cinematic video in one session.
A screen-recorded read-through of Anthropic's own prompting playbook for Opus 5.5, tip by tip, with a mid-video pitch for the creator's $9 Claude Code community.
A three-level walkthrough of prompting, saving a design system, and layering in AI-rendered images to turn one topic into a branded Instagram carousel.
00:08
00:17
00:29
00:34
00:52
01:09
01:21
01:34
01:43
01:52
01:58
02:06
02:14
02:24
02:42
02:47
02:52
03:04
03:15
03:30
03:48
03:54
04:00
04:06
04:12
04:20
04:28
04:36
04:52
04:59
05:11
05:16
05:43
05:49
06:06
06:24
06:27
06:40
06:50
07:09
07:21
07:35
07:43
07:48
08:09
08:18
08:26
08:33
08:39
08:43
08:55
09:03
09:12
09:20
09:28
09:40
09:43
09:58
10:10
10:19
10:29
10:39
10:42
15:25
08:27
09:05
07:20
07:27
09:27