Principles for Building with Agents

Ten operating principles for software work where AI agents do most of the implementation. Each one names a failure mode and the discipline that prevents it.

This is the philosophy layer. The skills in ~/src/skills are the operations; this file is the why. When a skill prescribes a behavior, it's usually one of these principles in action — referenced by name and number.

The 10

1. Implement to learn

The fastest way to understand a problem is to build the wrong solution and feel where it bends. Don't try to design the correct architecture upfront when the constraints are still fuzzy. Implement the obvious thing, run it against real cases, and let the failures teach you what the design actually needs.

Failure it prevents: paralysis-by-architecture, where weeks of diagrams get spent on a system whose real constraints only emerge under load.

How to apply: ship the smallest end-to-end version first (even if hacky), then refactor with knowledge you didn't have before. Use /missions between milestones to ask "what did we actually learn?" and update the spec.

Operationalized in: /grill-me (interrogate before building, but don't over-design), /missions (each milestone is a learning checkpoint), /vsdd (the spec evolves with implementation).