Generate skills for AI, so that it can work with Smelter.

[JBRS307]

Private repo with skills showcase: https://github.com/smelter-labs/smelter-ai-showcase

[wkozyra95]

Few things to consider:

• Only skills? Maybe something else too e.g. mcp
• Keep in mind that we will be adding something for our internal usage. Think about where we should keep those tools and how to separate them from skills for users.
• I tried to generate AGENTS.md in the past. https://github.com/software-mansion/smelter/blob/%40wkozyra95/generate-agents-md/AGENTS.md Don't use that approach, but maybe there is something useful there

[JBRS307]

Skills for ts-sdk will be added and it is currently work in progress. As for MCP, this is not a good use case for it.

[JBRS307]

As for what can be done, there are 2 possible approaches:

• Rewrite the docs as skill and add some explicit instructions (default use cases for components, code snipptet for WebRTC connection from browser etc.). More work with handling that, but more control over contents of the skill.
• Create a skill just with a SKILL.md file that will reference the actual Smelter docs. Works almost as good as the bundled version, however is more prone to hallucinations (e. g. Notoriously using nonexistent style prop in the InputStream component). This approach gives us less control over skill's content, but creates MUCH less overhead.

Token usage is nearly the same for both approaches.

I personally am in favour of the first approach. We can make our life easier by creating agent skill that will write this for us. It should base on 3 things:

• The skill-creator skill by anthropics. What is currently in the showcase repo was created with that (it was tweaked manually, but base was vibed)
• The previous skill for Smelter TypeScript SDK (if exists)
• The docs for Smelter TypeScript SDK

Please note that this still needs to be created.

[JBRS307]

About the skill distribution. The easiest way would be to host it in a repo in smelter-labs (maybe a .zip or .tar.gz file). Then we should add the llms.txt file to the smelter.dev, this file should contain information about the skill location and should suggest downloading it to the user, so the agent can work better with Smelter TypeScript SDK.

[JBRS307]

Now for the problem with llms reading our docs. The problem is not in the content, if the model is explicitly directed to specific doc pages everything works fine. However if the model starts to crawl the docs (for example after recommending Smelter to a dev) it is lost very quickly and starts throwing 404 http errors.