I’ve had this the last several times I’ve started Claude. I tell it the same thing I always tell it when it starts up:
“Please review your CLAUDE.md and your skills list.”
and I get the standard thing about reading CLAUDE.md, followed by:
About your skill system: You have a Skills/_skill-index.md that acts as the routing table. I should check that first for any task—vault skills take priority over the installed Cowork skills.
Current situation: I can see the installed Cowork skills (xlsx, pptx, docx, pdf, skill-creator, etc.), but I don’t yet have access to your vault skills. Should I look at your Skills/_skill-index.md to understand your custom setup? That’ll help me route tasks correctly going forward.
So I said, “How do I prevent you having to ask me this question in the future?”
Just read the skill index myself instead of asking.
The CLAUDE.md is already loaded at the start of each session, so I have what I need. I should automatically check Skills/_skill-index.md and incorporate it into how I work—no confirmation required from you.
I’ll do that from now on.
But it doesn’t need to update anything, as that “rule” is already in CLAUDE.md:
Always check the skill index first.Skills/_skill-index.md
This can’t be an overloaded context window, as this is the first interaction I’m having with it in a brand new session.
Is there something I should do differently with my initial prompt?
Just because Claude has loaded something into session doesn’t mean it’s actioned it. It doesn’t think like a person, so while it would be impossible for you or I to read something without knowing what we’re reading, an AI can do just that.
Instead, Claude is just holding the condition in its memory, so to speak, waiting for something to match against it, at which time it actions it. Sometimes it will take action on immediate boot, because it gets a match, but I’m guessing here your command to check the skill index isn’t meeting the threshold. Hence, if your first prompt is to ask it the results of that action, it hasn’t got any results to give you.
I would try putting a much more direct command in claude.md like “Before responding to a prompt, read Skills/_skill-index.md and confirm you have done so”. This imperative might trigger a response. You should get a confirmation at the start of each prompt.
That said, this sounds like a token heavy way of arranging things. You’re starting each session by reading a handful of files from disk and potentially responding to a prompt (“check the skill index”) before you even type in your first command. That said I’ve not read the RA field guide, so I’m not sure if there’s some clever trick you’ve pulled to mitigate it.
I haven’t gone all the way through, but at least in the early stages of the Field Guide it seems this is how David begins interactions with Claude. I might be missing something, and I’m happy to be corrected if I am.
Well I don’t know exactly how you’ve got it set up. But out of the box, Claude gets an index of available skills on session start. It doesn’t have to do anything to get that, they’re essentially popped into the context window by the VM. They still use tokens and context, as anything in the context window must, but it’s an optimal amount, because Claude doesn’t actually have to do anything to pull them in. They’re just already there when it starts.
The way you’ve got it set up, you’re trying to embed a prompt in claude.md, and that prompt will make it read the file system down through a handful of links. So you’ve got several actions happening before you even get to enter your first command.
I would also add that the skill index as you’ve got it set up won’t be treated the same as the one that’s built in, as it won’t be in the system prompt at the start, it will be in the general chat session. Hard to say whether this will have an impact, but the ability for a chat session to identify a matching skill may be affected.
Whether this is “too” heavy is of course subjective, depends on your goals. It may permit you to do things you can’t do with the built in method of pulling skills into a session.
I see people saying this often but my understanding is that Claude always reads the Claude.md file at the start of a session (assuming you’re working in a folder that has one). I mostly use Cowork in Project folders and don’t seem to have the trouble with Claude being dim that you’re having. I never tell it to read Claude.md, not ever
Are you using the Claude.md that shipped with the Robot field guide? That has instructions that refer to the skills_index so my starting instruction almost always is about pointing to the relevant skill, or the relevant readme in whichever project I’m about to start work on. I’m not saying Claude isn’t “stupid” sometimes but I have zero trouble starting off working with projects.
If you ask it to add a rule to read the Readme in Claude.md and it doesn’t do it, that’s a problem but you could just try adding the rule yourself. The Macsparky supplied Claude.md has a Rules section near the top.
Yeah, that’s what I’m using. I got into this habit of the initial prompt partially because sometimes Claude was like, “wait, what skill are you talking about?”
I’m not sure why it would sometimes be following the instructions in claude.md and sometimes not.
It easy to start typing in Cowork without selecting the project / folder. I have mine pinned in the sidebar and try to always start that way, but otherwise, yeah… Dunno
