@stultulo
it's a long complicated story (among, at least, one other long, complicated story)
but one result is: i've started enforcing skills and workflows as code, which is, like... ok i guess we could just call it code.
is how you get agents saying things like "The CLI expects an exact feedback line in the critique, and the artifact says “Verdict: accepted…” in prose but not the machine-readable `Decision: accepted` line."
or: "I’m running the phase verifier again to see whether README-7C needs another fresh review cycle or can now advance."
...like it just seems cleaner this way
but also: idk tho aha, I'm not really sure 🤔
i always just kinda do my own thing, but it seems like people are converging on this sort of approach,
e.g., The skill for when agents need to collaborate and write a spec is now but a thin dispatcher, as Codex likes to say, and reading the skill now, it really doesn't have to be more than this:
1 reply
1 recast
8 reactions