But it was not meant to be language agnostic anyways
webofbits
webofbits
0
webofbits
dimitarvp
That skill list is great. It does seem to me that it begs to become a tree or a DAG though. However, since this harness is meant to work with weaker models (I mean, everything is weaker than Fable and Astra, that’s what I mean) then it’s probably best we don’t hand those models complex multi-step instructions and have them go through these skills, manually or automatically, one by one, with human-in-the-loop inbetween; is that your intention?
0
webofbits
We’d have to manually ref them with /skill REF but I haven’t tested enough to see how it goes. Playing by ear. I’ll split into two commands for builtin and custom skills so that users dont get overwhelmened.
Curious to know how @mudasobwa is handling skills in his harness
0
mudasobwa
Creator of Cure
Please define “weaker” and provide a prooflink. I personally think that’s a marketing BS.
0
webofbits
webofbits
Probably weaker means models with less context and tools but if we pick sonnet model from ollama we prob just have to add harness to it? claude code harness isn’t open source unfortunately
1
dimitarvp
Everyone has their own definition, which is part of the problem nowadays, of course.
To me it means “can be left alone to work on a longer task without screwing up too much”. (Though they do still screw up, just not as much as others.)
In my professional work that translates to our agents having to pay attention to 100+ policies and coding practices and invariants / laws – Fable does much better there compared to Opus. I constantly had to remind Opus to look for shared project helpers and not introduce its own, never introduce loose financial assertions (only exact matching is allowed in our tests, never other comparison operators), check error response tuples from each 3rd party system and handle them (saga pattern) – three quick examples where Opus kept screwing up to the point of me almost giving up on using LLMs altogether. Happily Fable got out literal two days later and things went okay-ish.
Not so happily, it still does stupid crap, quite a lot, but on a higher level. That I can handle better, though barely.
What’s your definition, and can you back it up?
Mind you: 100 people agreeing with it means practically nothing. I mean, there are books out there telling us people must be killed if they wear mixed-fabric clothes after all. Surely millions of people agreeing with that book makes the rule absolute and correct. And: Spiderman exists in thousands of comics so yep, he’s real. Let’s have those nuances in mind when we argue.
0
mudasobwa
Creator of Cure
I don’t 
I don’t find skills per se to be of any good. Of course, I support /skills command, but mine is empty.
1