Autonomous Agent Design — Design skill for Claude Code

Autonomous Agent Design

Ship agents that stay on task instead of drifting halfway through.

Attention is a budget, and most agents blow it on the wrong tokens.
This is the toolbox for spending it well — and for catching
the misalignment patterns before they reach production.

Claude Code skill · context engineering over prompt engineering · system prompt altitude & U-curve ordering · proactive MANDATE pattern (per Anthropic's Agentic Misalignment findings) · anti-hallucination grounding · just-in-time retrieval & compaction · tool design

It condenses the best of agentic context engineering & prompt engineering research into one loadable skill:

• Mental model — context engineering > prompt engineering; attention is a budget
• System prompt altitude — strong heuristics + decision margin, U-curve ordering (anti "lost in the middle")
• Proactive MANDATE pattern — urgency directed at the goal/user, never at the agent's own survival (per Anthropic's Agentic Misalignment findings), with anti-sycophancy rules
• Anti-hallucination / grounding — no claim without a tool call, cite real artifacts
• Context window optimization — just-in-time retrieval, compaction, external memory, prefix cache, subagents
• Tool design — few tools,