I am interested in an analysis for my most complex libraries if you have the time : jsv, oaskit and gen_mcp.
Regarding skills as tasks vs knowledge to me it makes no difference, at the technical level the agent will match a topic and ingest the related text, whatever that is. A “how do I do X” task description is still knowledge. Promps that you invoke manually I call that “slash commands” but once it’s in the context, it does not make much difference, except that in VS Code copilot the path to the file is prefixed with “follow the instructions in” which directs the LLM to explicitly follow what’s inside instead of just getting some knowledge, it’s more suited for tasks in my opinion.
Also LLMs are not Elixir specialists and will follow the most used patterns all languages wide. I stopped to tell them to write tests like this:
assert %{
foo: 123,
bar: "hello"
} = stuff()
because they just don’t listen. The want to write this as:
assert x = stuff()
assert x.foo == 123
assert x.bar == "hello"
This is just what typescript and python does. I’m now willing to try to write a “testing with elixir” skill to see if this can be improved.
@SyntaxSorcerer this pattern is interesting! I’m using SDD tools like openspec sometimes, but I always find those frameworks a bit too rigid. Thanks to the nature of LLMs it’s always possible to steer away from the basic workflow, but I believe for smaller tasks it could be more flexible to have a “handoff to impl”, “handoff to to fix”, “handoff to tests” kind of prompts.
Ah but saying that I realize that I’m willing to adapt AI-as-a-tool to my natural workflow (which is mostly TDD). And that there are two paradigms that are actually clashing: Do you use the AI as support and a tool to generate parts of the code faster, or do you want it to be an independent developer working on its side and have a coworker relationship with it. Or is it a spectrum between the “10x senior” and the “vibe coder”? I think that if you want an autonomous agent you have to fully embrace it (which I’m not ready to do) and provide many skills to describe the “AI developer” as a profession on top of the “competent programmer” skills.
Well sorry this was mostly off-topic 
