跪拜 Guibai
← Back to the summary

Claude Skills Pipeline Slashes Short-Video Monthly Bill From $86 to $16

I used Claude Skills to cut my overseas video monthly bill from $86 to $16

The most off-putting part of making short videos for overseas markets is the generative video models: a few tens of dollars of credits per month simply don't last, and the character frequently breaks or switches between shots. Changing one line of dialogue means discarding the roll and starting over. Anthropic just donated Claude Skills to the Linux Foundation. This open standard, already integrated by more than 70 platforms, solves the old problem of AI having no process. I used it to hand-roll a pure-code stickman animation pipeline locally. A single 15-second 60fps clip measured just $0.54, and editing dialogue costs zero re-rolls.

Before you follow along, check your computer's prerequisites: install Node.js, FFmpeg, and Chrome or Edge. The export script will call them directly.

Open Windows PowerShell, enter the working directory and start Claude Code:

claude

After startup, you'll see the version number and a waiting cursor in the terminal:

Opening Windows PowerShell, starting Claude Code, initial welcome screen

Below I'll walk you through step by step.

One: First look at the negative example: what happens without a Skill

Let's see what happens if you let the model write raw without skeletal constraints.

Paste the first prompt into the input box. It first checks the current directory for an existing Skill, then writes a webpage with no Skill:

First help me check the .claude\skills folder under C:\Users\your username\OneDrive\Documents\字流AI实测 to see if there's a stickman animation Skill (only look at this directory, don't scan the rest of the computer). Then, without using any Skill, write a single-page 15-second stickman explanation animation webpage at C:\Users\your username\OneDrive\Documents\字流AI实测\stickman-demo\v1-no-skill.html, about "the pain points of a one-person company making short videos for overseas", with 3 storyboards, horizontal 16:9 for desktop, no irrelevant decoration. When done, tell me the result and time.

After about 2 minutes, the terminal confirms there is no stickman Skill in the current directory, and it took 64 seconds to write an 11.6 KB single-file webpage v1-no-skill.html.

Terminal confirms no Skill and generates a single-file animation webpage v1-no-skill.html

Double-click open the webpage in Google Chrome to play it.

Cards appear for topic selection, shooting, editing, etc., but the stickman's head-to-body ratio is stiff, the arm lines are skewed, and the entire page's animation logic and HTML are all mixed together.

Browser shows raw stickman animation first act and rough lines

This is the common disease of no specification: the large model improvises each time, the skeleton proportion has no mathematical baseline. As soon as you change a shot or pose, the figure immediately distorts.

Two: Hand-craft the stickman Skill: package skeleton specification and export tool

To keep the character motion from ever breaking, the skeleton baseline and motion rules must be packaged into a Skill.

Back at the Claude Code terminal, paste the second prompt:

Now help me create a Claude Skill for making 60fps stickman explanation shorts at C:\Users\your username\OneDrive\Documents\字流AI实测\.claude\skills\stickman-animator. Besides writing the SKILL.md specification, package two script files: one is a rendering library engine.js that fixes the protagonist's skeleton proportions (a stickman programmer with black-rimmed glasses) and smooth joint actions; the other is an export script export-mp4.js that uses the local existing Edge or Chrome browser together with lightweight tools to quickly record a webpage animation into MP4 within 1 minute (to avoid timing out by downloading a large browser). When done, tell me which files are in the folder.

Waiting about 8 minutes, Claude Code finished writing the specification files and self-checking, and configured 11 preset poses including wave, point, think, panic, etc.

Terminal shows stickman Skill's preset poses and 52-second export test result

To confirm the character skeleton won't deform, send the third prompt to view core implementation code:

Print the SKILL.md content and the core character skeleton code in engine.js on the screen for me.

After about 35 seconds, the terminal prints the core function code responsible for computing the polyline of the torso and limbs.

Terminal outputs fixed bone length RIG configuration and single-joint angle interpolation core source code

Its underlying logic is very clean: the stickman's torso and limbs are essentially a polyline, and each joint is controlled by a single angle. Every frame is computed through single-joint angle interpolation, and coordinates are calculated strictly according to the fixed bone lengths in the RIG.

As long as the bone lengths and the relative parameters of the black-rimmed glasses are locked, no matter what actions the character makes or how the shot changes, the facial features and body proportions won't deform.

Three: New conversation: one sentence automatically triggers the Skill to produce a complete video

Now that the Skill is built, how does it actually work?

Enter /exit in Claude Code to close the previous conversation, then type claude again to start a fresh clean session.

Paste the fourth prompt directly:

Help me create a 15-second 16:9 horizontal stickman explanation animation at C:\Users\your username\OneDrive\Documents\字流AI实测\stickman-demo\v2-skill.html, theme "one-person company making short videos: from not being able to afford expensive generative video credits, to using pure code to reliably run 60fps animation", with 3 acts (1. worrying over the bill at night 2. character face-breaking and shuttling, shaking head 3. using code to smoothly generate, cheering happily), with bilingual subtitles at the bottom, and a real-time FPS display in the top right corner. Directly call the engine.js rendering library in our stickman Skill to draw it. Don't add irrelevant background decoration. When done, tell me the time and how you triggered the Skill.

In the new conversation I did not specify any complex skeleton parameters. Claude recognized the keyword "stickman explanation animation" and automatically invoked the stickman-animator Skill just built. If your environment's model doesn't automatically recognize it, you can also type /stickman-animator in the terminal to force invoke it.

Here's a small pitfall I hit in practice: If quota runs out or a countdown appears during generation, don't panic, and don't close the window to restart. Stay in the original terminal and just type "continue", and it will finish v2-skill.html smoothly. The terminal showed that generation plus self-check took only about 1 and a half minutes.

Terminal shows automatic invocation of stickman Skill while making animation

Open the generated v2-skill.html in Chrome and play it full screen:

The stickman wears black-rimmed glasses, and the body proportions are natural and coordinated. In the top right corner, a green "60 FPS" runs steadily, and bilingual subtitles scroll at the bottom.

Browser plays the first act of the final video: protagonist with glasses keeps stable shape and shows 60 FPS

Four: Change dialogue and add effects: editing code costs no re-rolls

When you run a one-person company going overseas, you have to distribute content to many platforms every day. The worst thing is that changing one word destroys the whole video. Traditional generative video models, once a shot is unsatisfactory, force you to re-roll, which costs time and credits.

Code-driven animation is extremely light to edit. Back at the terminal, paste the fifth prompt to add two effects:

I just opened v2-skill.html and looked. Overall it's stable, but I want two changes: first, add a prominent red "-$$$" floating text effect above the computer screen in act 2, to emphasize costs burning money; second, add a green "60FPS / $0" highlight badge next to the stickman when he cheers in act 3. Update v2-skill.html directly, and tell me what you changed when done.

After about 1.5 minutes, Claude adjusted the code for the corresponding frames, updated the webpage directly, and noted in the terminal that it also re-exported the MP4 video.

Send two modification requests to Claude

Refresh v2-skill.html in the browser and look at the modified frames:

In act 2, a red "-$$$" effect rises above the computer screen; in act 3, a green "60FPS / $0" highlight badge clearly lights up beside the stickman. The protagonist's black-rimmed glasses and skeleton were not affected by the changes.

Browser checks act 3: green 60FPS and $0 badge added beside stickman

Five: Local rapid final video: 52 seconds to export 60fps MP4

After the webpage is tuned, short-video platforms need a real MP4.

We directly call the export script packaged in the Skill. Paste the sixth prompt in the terminal:

Now directly run the export-mp4.js export script in the stickman Skill. Record the just-updated C:\Users\your username\OneDrive\Documents\字流AI实测\stickman-demo\v2-skill.html into a 15-second MP4 video file, save it to C:\Users\your username\OneDrive\Documents\字流AI实测\stickman-demo\final-stickman.mp4. When done, tell me the file size and time.

It directly calls the locally installed browser for headless rendering, and FFmpeg does frame-by-frame encoding. The whole process uses no cloud GPU compute.

It took just 52.2 seconds. The terminal shows final-stickman.mp4 generated successfully, file size only 0.81 MB, standard 1280x720, 60fps, exactly 900 frames.

Terminal shows export-mp4.js only took 52 seconds to export high-definition video

Open final-stickman.mp4 with the system player. The three acts transition smoothly, bilingual subtitles are clear, and you can directly drag it into editing software or publish it to social platforms.

System local player smoothly plays the exported 15 second high-definition short clip

The final output looks like this:

Tested final video final-stickman.mp4 (15 seconds)

Six: Compare costs: How much did the real bill actually save?

Costs cannot be estimated by gut; you must look at actual spend.

Independent developers are sensitive to every cent. I initially assumed a single clip was about $0.05, but on reflection the number had no basis. I went back to the terminal, pulled up the billing command, and decided to check properly:

/cost

The terminal pulled up the real session bill:

Terminal shows session bill: Haiku and Sonnet real total spend $0.54

As a comparison, open Runway's official help documentation in a browser and look at the credit conversion and generation rates for various models.

Runway official help doc shows compute unit prices for each model video generation

Back in the terminal, paste the seventh prompt to generate a comparison webpage based on the real bill, and update the actual spending into it:

Help me look up the current official pricing for Runway Gen-3 and Luma Dream Machine to generate a 15-second video, and combine it with the actual Claude Token bill from this conversation that generated and exported this 15-second stickman animation (first help me estimate or read the current session's actual Token consumption in dollars, and cross-check it with what I saw in /cost). Create a clean dark cost-comparison webpage at C:\Users\your username\OneDrive\Documents\字流AI实测\stickman-demo\cost-compare.html, comparing the cost per 15-second video, monthly cost for 30 videos, character consistency, and rework cost to change one line of dialogue. Tell me the results when done.

Claude wrote the comparison page and checked and corrected the real bill to $0.54.

Terminal confirms writing the measured cost $0.54 into comparison page cost-compare.html

Open cost-compare.html in the browser and all data is clear at a glance:

Browser full screen shows cost and feature comparison between pure-code stickman and mainstream generative video models

Pulling out this plain-language comparison table makes the difference even more obvious:

Approach Per 15-sec cost Monthly for 30 Change one line of dialogue cost Character consistency
Claude Skills code pipeline $0.54 (measured total session) $16.20 $0 (change code, re-export 52 sec) Extremely stable (fix skeleton)
Runway Gen-3 $0.75 $22.50 Pay to re-roll again (+$0.75) Prone to shot drift
Luma Ray3 ~$2.88 $86.40 Pay to re-roll again (+$2.88) Prone to face breaking and switching

Generative video models require paying again for a re-roll to change one line of dialogue; the pure-code approach changes two lines of code, re-exports locally in 52 seconds, and the character skeleton never breaks.

How to choose and my suggestion

Going through this entire process, my judgment is clear:

This approach fits overseas content creators and one-person company independent developers who have limited budgets, need high-frequency explanatory short videos, and suffer from generative video models that frequently break faces and switch characters. It turns an uncertain "re-roll" process into "deterministic production" driven by a fixed skeleton engine, eliminating cloud GPU rendering costs and solving the old problem of character deformation across shots.

But it has boundaries: it is not suitable for those pursuing photorealistic effects, complex cinematic camera moves and effects, or non-technical users whose computers don't have Node.js, FFmpeg, and a basic development environment.

If you are also being tormented by generative video models' credit burn and re-rolls, follow the steps above and build your own animation Skill locally.


👋 I'm Meng Jian, ex-Tencent T11, ex-ByteDance tech leader, now doing AI programming full-time.

🔥 More AI programming practice:

Found this useful? Like + bookmark is the biggest support 🙏

This article was organized with AI assistance, and the content was tested and reviewed by the author.

Comments

Top 1 of 2 from juejin.cn, machine-translated. The original thread is authoritative.

哦豁没了

I thought it was a manhua drama....

孟健AI编程

Manhua dramas are indeed everyone's first choice for going overseas; this stickman animation style leans more toward logical explanation.