Topics

Customization

Skills, hooks, subagents, MCP, plugins, and AGENTS.md — ways to extend Claude Code for work you repeat.

A stack of instruction slips rises past a low ceiling bar, with four cards failing in different ways.

YouTube

Is the 200-instruction skills ceiling still real?

Laurie Voss traces the ~200-instruction “agent starts forgetting” figure to a 2025 aside from Dexter Horthy, then reruns an IFScale-style keyword-inclusion test. Last year’s remaining models still fell apart around 200–300 named constraints. The then-current frontier, in his run, walked through 500 required words and held until roughly 2,000 (best around 5,000). That is his claimed ~10x jump, not a product spec and not a Radar rerun. Longer skills files are now plausible; the live problem is that models fail in four different ways, and a plausible answer can still be incomplete.