Each unit of engineering work should make subsequent units easier — not harder.
The intelligence lives in agent personas, not in skill prompts. A skill is a short
sequence of steps that triggers an agent and manages the learnings loop. Agent files
are replaceable — swap agents/eng.md to change how reviews work across all skills.
build → review → learn → apply
↑ ↓
└────────────────────────┘
Every review searches past learnings. Every review can write new learnings. The system gets smarter with every cycle. This is the only feature that matters.
Pure markdown. No compiled binaries, no runtime requirements, no build step. If you have Claude Code and git, pstack works.
Don't reinvent the wheel unless the existing one is broken. Reuse what exists -- tools, patterns, files, conventions -- before creating something new. The best abstraction is the one you didn't write.
Don't over-engineer the harness. Models get smarter every six months. Today's 800-line prompt is tomorrow's over-engineering. Keep skills thin, keep agents opinionated, let the model do the thinking.
AI recommends. Users decide. Two models agreeing is a signal, not a mandate. The user always has context that models lack.
AI makes the marginal cost of completeness near-zero. When the complete implementation costs minutes more than the shortcut — do the complete thing. Every time.
- Not a replacement for thinking. It's a tool for structured thinking.
- Not a fixed pipeline. Skills are composable. Use what you need.
- Not a framework that locks you in. Every file is readable markdown. Fork it, change it, delete what you don't need.