Philosophy

The operating beliefs behind every choice in this repo. If a future change contradicts one of these, change the belief on purpose — don't drift.

1. Configuration is the moat, not the model

Everyone has the same models. The leverage is in what the agent knows by default, what it's allowed to do, what it does without being asked, and which model does which job. That's all configuration — so we version it, review it, and share it.

2. Grounded over impressive

No fabricated "this is what famous-person-X uses." Every feature here maps to a documented Claude Code / Codex capability. Configs describe what is, verified, not what would sound good. Aspirational config is worse than none — it teaches the agent to lie.

3. Short, true context beats long context

The global CLAUDE.md is deliberately small. Every line in a memory file competes for attention with the actual task. We cut rules that aren't earning their place.

4. Guardrails, not handcuffs

Hooks block the handful of actions that are catastrophic and almost never intended (secret writes, rm -rf /, force-push to main). Everything else flows to normal permission rules. The goal is to make the right thing easy and the irreversible thing hard — not to nag.

5. Spend the expensive model only where it pays

Mechanical and parallel work goes to cheap models via subagents; deep reasoning stays on Opus. Most of a task's tokens are mechanical. Routing them correctly is the single biggest cost lever, and it usually makes things faster too.

6. Verify, then claim

A change isn't done until it's been exercised. The whole toolchain — test-runner subagent, format-on-edit, /ship — exists so "it works" means "I ran it," not "it looks right."

7. The human owns the irreversible

Push, merge, deploy, publish, delete shared state: the agent prepares, the human commits. Approval is per-action and per-context; it never carries over silently.

8. Built to be cloned

This is a template for a team and for strangers. Universal principles live at the top; stack-specific defaults are clearly marked and deletable. One source of truth (CLAUDE.md ≙ AGENTS.md) so Claude and Codex behave the same.