Anthropic says it cut Claude Code's system prompt by over 80% for newer models, partly by loading skills only when needed. Does your own setup work that way?
retained text
Anthropic's Applied AI team published an article on Oct 22 about folding subagents back into one main agent. It is aimed at enterprise teams building on Claude Platform, but a few lines apply to anyone who sets up Claude with skills. The line that caught my eye is that they reduced Claude Code's system prompt by over 80% for more advanced models, and that part of the cut came from loading skills and tools only when they are needed. Their broader claim is that across several enterprise deployments, one agent with skills frequently beat both a giant single prompt and a subagent setup on quality, often at lower cost and latency. The one hard number is a roughly 2x faster runtime for a customer they do not name, so I would treat it as directional. They also say subagents still make sense for parallel work, for measurable context bloat, or when one agent keeps picking the wrong tool. What I cannot tell from the article is how this plays out for individual users instead of enterprise fleets. So I am curious: * Do you keep most of your instructions in [CLAUDE.md](http://claude.md/) or project instructions, or push them into skills that load on demand? * Have you moved something from a subagent into a skill, or the other way around, and noticed a difference? Source is in the first comment.
document fetched 2026-10-30 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1xidh5x/anthropic_says_it_cut_claude_codes_system_prompt