Agentic Coding

Anthropic finally documented the skills system end-to-end.

A 33-page Claude Skills guide walks through setup, configuration, and best practices for the skill architecture you've been building manually [Source: Facebook]. This is the official playbook—not community-sourced patterns or reverse-engineered approaches. Combined with the AAS Core catalog and PM skills library from recent days, your skill stack now has canonical documentation backing it. The guide covers ground that took most of us months to piece together from scattered threads.

Save this one—it's the reference you'll keep returning to.

Your agent might be burning half its reasoning budget before doing real work.

A new analysis argues most AI coding agents waste roughly 50% of their context window on redundant information, outdated context, and poor memory management [Source: Medium]. This connects directly to why tools like Graft—which you saw two days ago slashing token use by 42%—matter so much. The fix isn't bigger context windows; it's smarter context loading. Your AGENTS.md structure and skill pre-flight checks already help, but the article suggests most setups still have significant waste to cut.

Check where your tokens actually go.