This would be much more interesting if it detailed how the prompt/skills were reverse-engineered. As it is it seems like this could just be the output from “hey Claude write me a system prompt that works like Claude Design”.
They have never attempted to hide their system prompts in fact they explicitly publish them for Claude (https://platform.claude.com/docs/en/release-notes/system-pro...) and Claude Code they provide a proxy option which then makes it trivial to see the full requests and responses including the prompts and tool usage...
honestly, i think you can just look at the network tab and see the "content" of the skills. Same has been true for their excel addin and bunch of other things.
I'm calling BS, sorry. It looks light, and barely anything beyond surface level of what we could all could guess would be in a system prompt. This smells nothing more of a "claude give a system prompt that anthropic would use as a system prompt for claude"
From what we know, there are some very specific details baked into the prompt as safety guards, where are those? Again calling BS and I'm not gonna waste more thought/words on this
In general I agree with you, especially about how things should work, and find the current trend of claiming LLM-washing code hugely problematic.
I do wonder what the analysis would be of these prompts were vibe-coded, though, since in general LLM output can’t be copyrighted without significant human authorship, and Anthropic is pretty noisy about minimal human supervision.
The minimal human supervision for prompts seems like a pretty silly take. LLMs are still pretty bad at creating good prompts for LLMs. Give it a benchmark and some feedback and it can brute force it, but far less effectively. And I thought the point of using LLMs for development was increased efficiency.
I can't even tell if this repository is based on prompts extracted from Claude Design or if the author had an LLM create all of these prompts in it from scratch.
The fact that they encourage and accept PRs indicates that this isn't intended as a direct prompt extraction exposure project - plus the license, which should indicate they have the authorship necessary to license that content.
Assuming this IS a complete ground-up implementation it really needs to link to demonstrations that it works. Without any evidence it's hard to justify spending time exploring it.
If you ask Claude Design itself to list the names of the skills available to it you get:
Animated video
Interactive prototype
Make a deck
Make a doc
Make tweakable
Claude API in prototypes
Frontend design
Wireframe
Export as PPTX (editable)
Export as PPTX (screenshots)
Create design system
Save as PDF
Save as standalone HTML
Send to Canva
Handoff to Claude Code
Which does not match the structure of this project at all.
I've been using Claude Design to make animated SVGs, and I've learned a thing or two about its limits and how to get around them.
One thing I've learned is that you have to ask it to first come up with a robust way to define the geometry and then apply that to an SVG. Without that first step, it just guesses at where everything should be that isn't directly connected with a node, and it is hilariously bad. But with that first step it is capable of creating some incredible geometry algorithmically from detailed instructions.
The other thing is that whatever tool prepares the svg for export will strip the animations as part of a sanitization process, it won't even see that has occurred. You have to ask it to export to a different file type like my-animation.svg.txt, and then obviously you want to inspect it carefully because svg can carry many exploits not related to animation.
25 comments
[ 0.17 ms ] story [ 35.2 ms ] threadThat would enable Anthropic to block the technique.
Side note: ironic use of an llm writing the readme.
From what we know, there are some very specific details baked into the prompt as safety guards, where are those? Again calling BS and I'm not gonna waste more thought/words on this
I don't think that is how copyright licensing works.
Seems pretty obvious to me.
I do wonder what the analysis would be of these prompts were vibe-coded, though, since in general LLM output can’t be copyrighted without significant human authorship, and Anthropic is pretty noisy about minimal human supervision.
> “I don’t prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do. My job is to write loops.”
> “Can confirm Claude Code is 100% written by Claude Code”
Both from head of Claude Code. Taken either literally or seriously they point to Anthropic shipping code with limited human authorship.
The fact that they encourage and accept PRs indicates that this isn't intended as a direct prompt extraction exposure project - plus the license, which should indicate they have the authorship necessary to license that content.
Assuming this IS a complete ground-up implementation it really needs to link to demonstrations that it works. Without any evidence it's hard to justify spending time exploring it.
It’s different than this one shared by the op, but Anthropic maybe updated the prompt
One thing I've learned is that you have to ask it to first come up with a robust way to define the geometry and then apply that to an SVG. Without that first step, it just guesses at where everything should be that isn't directly connected with a node, and it is hilariously bad. But with that first step it is capable of creating some incredible geometry algorithmically from detailed instructions.
The other thing is that whatever tool prepares the svg for export will strip the animations as part of a sanitization process, it won't even see that has occurred. You have to ask it to export to a different file type like my-animation.svg.txt, and then obviously you want to inspect it carefully because svg can carry many exploits not related to animation.