docs: add evidence-based agent prompt principles

[mpawliszyn]

Summary

• 12 evidence-based principles for writing effective agent prompts in Fowlcon
• Derived from comprehensive research across cognitive science, LLM behavioral studies, multi-agent orchestration patterns, and production agent tools
• Will directly guide implementation of agent prompts in Phases 2-4

Key Principles

1. Agents are tools, not peers -- structured output, no back-and-forth
2. Format constraints control behavior -- Aider showed 3x improvement
3. Negative constraints before positive -- the RPI documentarian mandate
4. Name the rationalizations -- Superpowers red flags table pattern
5. Mechanical verification over self-assessment -- scripts, not self-reports
6. One agent, one job -- tool restrictions enforce specialization
7. Context inline, never file references -- guaranteed visibility
8. Tree structure limits -- 2-3 levels, 3-5 children/node (cognitive load research)
9. Unified diff format -- omit hunk line numbers, include context lines
10. Design for fresh starts -- 39% multi-turn degradation, state on disk
11. Supervisor mode for parallel agents -- failures as data, not exceptions
12. The reviewer is the protagonist -- anti-automation-bias design

Research backing

Backed by 16 research documents (6,800+ lines) covering multi-agent patterns, cognitive science, production tools, and benchmarks. Full research library in the private design repo.

Test plan

• Read through all 12 principles for clarity and correctness
• Verify research citations match source material
• Confirm principles align with existing AGENTS.md conventions
• Check for conflicts with the 10 core principles in README.md

🤖 Generated with Claude Code