Daily AI Catchup
A folder of instructions, not a new model
A "skill" is not a different, smarter AI switched on for the occasion. It's a written playbook the same assistant reads when a task matches it: how to fill out a specific kind of PDF form, how to build a slide deck, how to follow your team's code-review checklist step by step. Sometimes the skill ships with a script or a template file it's meant to use along the way.
The skill itself doesn't do anything on its own. It's the instructions the assistant follows once it decides this is the right playbook for what you just asked.
The context-window problem skills are built to dodge
An assistant with dozens of skills has a genuine problem: a context window is finite, and every skill's full instructions take up real space in it. If all of them sat in context on every single turn, a handful of skills would eat the whole window before the conversation even started — long before you'd typed your actual request.
This is the same "lost in the middle" problem a crowded chat has, one layer up. The fix isn't a bigger window. It's not loading what isn't needed yet.
How the loading actually works
Instead of holding every skill's full instructions at once, the assistant keeps just a name and a one-line description for each skill it has available — a tiny index, not the contents. When your request comes in, it matches against those short descriptions to find the one skill built for this, and only then loads that skill's full instructions into context for the first time.
The other skills haven't cost a token. Their instructions are still sitting unloaded, exactly where they were.
The agent doesn't decide what to forget. It decides what to load in the first place.
What this means in practice
Because matching happens against that short description alone, a vague or generic one makes a skill genuinely hard to find later — the full instructions inside it don't matter if the one-liner never gets it matched. And any bundled scripts or templates a skill ships with are just files sitting there until the loaded instructions tell the assistant to actually use them.
Updated August 2026 · 2 min read