Posted by domhudson 20 hours ago
This is true. Transformer has several orders of magnitude more working memory than any human. Compared to transformer we all have executive dysfunction.
By default they explain things assuming I have infinite processing bandwidth. I do not! I have several zeroes less than they do.
But this seems to focus very much on telling the user what to do, whereas usually I'm telling Claude what to do. It feels like this inverts the relationship and wants to turn me into a reverse centaur.
Wouldn't "lead with the answer" be better than "lead with the action"? But sometimes answers do require detail and explanation. I just want to get rid of all the unnecessary prose.
And stop using Opus 9 Pro Max XHigh 10.0 for everything. If you choose a hyper-thinking model for asking the weather you can’t but expect yapping.
I also don't know how much to trust the model, but I've had the model tell me specifically that certain aspects of ASD-STE100 are unactionable and will just create more noise.
The OP's own skill even leads with something in a very similar vein:
> These rules apply to every response for the rest of the session, not only this one. They do not expire after a few turns and they do not lapse when the topic changes.
My understanding is that phrases like this are at best a _very_ weak signal to the model. It's simply contradictory to how the model works at a level that can't be overridden by injecting tokens.
I placed those instructions in setting > personal instructions, in my global CLAUDE.md, and in each project’s CLAUDE.md. I also use the concise output style in CC. If I choose a high thinking model it will start deviating in long sessions, then I just remind it in my next message:
“Remember ASD-STE100 style.
[rest of my message]”
BLUF sticks a lot easier than STE to be honest… but Claude knows what STE is, and using the “ASD-STE100 style” locution avoids the compliance issue (it’s true that strict ASD-STE100 compliance isn’t really possible, nor desired)
https://gist.github.com/hbbio/2faf096cbb77e197233ab9a2958beb...
https://news.ycombinator.com/item?id=46871173
Anyways the most layman way I’ve seen it explained is this: skills help save token usage for the right context. Not every request needs all instructions all the time - running tests is different than reviewing a PR, so why should the context window have instructions for both on every request?
So now you split instructions into “skill” files, which are basically opinionated markdown files. And you invoke those with something like /grill-me in the prompt depending on what you’re doing.
There are some steps to have the agent automatically know what to invoke for you but in my experience this automation is hit or miss.
It is also challenging to keep track of a growing library of skills and keeping those up to date.
So YMMV regarding skills. I typically keep things in a single markdown file even if the context window gets a bit bloated.
My point is that this sounds more like a general AGENTS.md use case, similar to defining tone of voice, output format, etc.
It just seems skills is the only way to distribute certain "behavior" as of today. But not everything is a skill IMHO, and not this is not one of them.
On the other hand, maybe we're seeing an evolution of what skills are becoming.