docs: agent prompt design guide

[mpawliszyn]

Summary

• Adds docs/guides/agent-prompt-design.md -- a practical guide for writing and maintaining Fowlcon's agent prompts
• Captures 10 design principles with evidence and rationale
• Includes how-to sections for writing worker prompts, handling external dependencies, and the three-tier hierarchy
• Decision log documents key choices and when to revisit them

Context

This guide was developed during Task 7 (worker agent prompts) design phase. Before writing the actual prompts, we worked through 8 open design questions covering tool access, version tracking, behavioral controls, output format, tone, and change boundary tracing. Each decision is grounded in analysis of published patterns from Anthropic, OpenAI, LangChain, Aider, SWE-agent, and others, plus instrumented experiments running research agents against real codebases.

Key findings that shaped the guide:

• Format constraints are the strongest measured quality lever (Aider's 3x laziness reduction)
• Mechanical tool restriction beats prompt-based restriction for role enforcement
• Workers are binary on external code (excellent locally, useless remotely) -- routing is a researcher concern
• Natural-language reasoning pauses work across models; model-specific keywords are brittle

Test plan

• Read the guide -- are principles clear and actionable?
• Check decision log -- does the rationale make sense?
• Verify no internal references leaked (should be all public information)
• Confirm guide is consistent with AGENTS.md conventions