本文由 AI 分析生成
建立時間： 2025-12-13
Summary
EN: Armin Ronacher (creator of Flask) argues that SKILL.md-based “Skills” (plain text instructions + bash tools) outperform MCP (Model Context Protocol) servers for AI agent tooling. MCPs consume 8K+ tokens in tool definitions, have an unstable protocol, and create a large surface area for failure. Self-authored bash-based tools with deferred loading are more token-efficient, predictable, and easier to debug. Skills that summarize their own capabilities work better than comprehensive documentation.
ZH: Flask 創始人 Armin Ronacher 主張，基於 SKILL.md 的「技能」（純文字指令 + bash 工具）在 AI 代理工具方面優於 MCP 伺服器。MCP 消耗 8K+ tokens 的工具定義、協議不穩定、失敗面廣；相比之下，自編寫的 bash 工具配合延遲載入更節省 token、更可預測、更易除錯。能自我摘要能力的 Skills 效果優於全面性文件。
Key Points
- MCP problems: 8K+ token definitions per server, unstable protocol (breaking changes), each tool call requires network round-trip
- Skills approach: SKILL.md file with plain text description of capabilities + bash scripts for actual functionality
- Deferred loading: tools are described in summaries; only loaded when needed — prevents upfront token explosion
- Self-authored tools are preferred: author controls behavior, no external service dependencies
- Focused 2-3 module skills beat comprehensive documentation (echoes SkillsBench findings)
- The MCP protocol is evolving fast — building on it means constant updates
Insights
- Ronacher’s critique of MCP token consumption is quantified and damning: 8K tokens per server means a moderately equipped agent spends its entire context budget on tool definitions before doing any work
- The “self-authored” preference is about trust and debuggability, not capability — you can reason about bash scripts far more easily than opaque MCP server behavior
- Deferred loading is a key architectural insight: describe tools cheaply (summaries), load them expensively (full definitions) only when needed
Connections
- SkillsBench paper in this vault: empirically validates Ronacher’s intuition — curated focused Skills +16.2pp, comprehensive docs negligible gain
- The Superpowers plugin article describes exactly this SKILL.md architecture in practice
- Claude prompt library: the deferred tool loading pattern aligns with Anthropic’s parallel tool call recommendations
Raw Excerpt
“An MCP server can consume 8,000+ tokens just in tool definitions before your agent has done anything useful. SKILL.md files with bash tools are text files — they cost almost nothing. Defer loading the full tool definitions until you actually need them, and your agent has a context budget to actually work with.”