I've run into the same issue and I still end up manually curtailing what's exposed to the model, limiting to the task at hand, but I like the idea of another (smaller I hope) model doing 70% of the clipping instead, automagically.
How? Using the agent SDK or Claude Code? If the latter, it'd be nice if they figured that out. There's a huge amount of quality of life things missing from Claude Code. It's a pretty raw frontend to the backend models. And either Claude Code or the backend models get convinced they don't need skills they've been asked to read or even built-in capabilities like reading PDFs.
You know what, I checked Opus 4.8's instructions to a review subagent the other day and it literally opened with
> You are a senior infrastructure/security engineer doing a thorough, adversarial code review...
I didn't say anything like that myself.
As a personal anecdote, I find that a lot of big prompts and skills use up context window budget and in many cases agents will eagerly try to use a skill even if it isn't super relevant or necessary for the current task. So when I have too many skills I have to spend a bunch of time toggling the checkboxes to figure out which ones are needed for the task at hand before starting...