🔄 Closed Loop Development

An agent skill that makes AI coding agents actually verify their work.

Edit → Verify → Fix → Verify again → Commit. No exceptions. No excuses.

Inspired by Peter Steinberger's battle-tested AI development workflow.

---

🤔 The Problem

AI agents love to write code and immediately declare victory. They skip tests, ignore linters, and commit broken builds. When called out, they rationalize: "it's a small change", "I'll run tests later", "the CI will catch it."

This skill fixes that.

✨ What It Does

🔍 Enter project  →  Discover verification stack
📐 Before task    →  Assess blast radius
✅ After edit     →  Verify (typecheck → lint → test → build)
🔧 If fail        →  Fix root cause → Re-verify (max 3×)
💾 If green       →  Atomic commit
🛑 If stuck       →  Stop → Report to user

🎯 Key Features

	Feature	What it does
🔍	Auto-Discovery	Detects verification commands from package.json, go.mod, Cargo.toml, CI configs, etc.
📝	AGENTS.md Generation	Creates project docs for future agent sessions when none exist
💥	Blast Radius Assessment	Estimates change scope → picks commit strategy accordingly
🔁	The Loop	Runs full verification (typecheck + lint + test + build) after every edit
🛡️	Authority Resistance	Pushes back when told to skip verification — even under pressure
🔧	Toolchain Bootstrap	Installs missing tools instead of skipping verification

🌍 Language Agnostic

Works with any language or project type:

TypeScript · Go · Rust · Python · Java · Swift · and more

The skill discovers the verification stack from project config files — zero hardcoded assumptions.

🧪 TDD-Tested (Yes, Really)

Built using RED-GREEN-REFACTOR methodology for agent skills:

• 🔴 RED — 9 scenarios where agents failed without the skill
  • Obeyed "skip tests" from authority without pushback
  • Bypassed broken toolchains entirely
  • Forgot to run tests after typecheck + build
  • Skipped AGENTS.md generation under task pressure
  • Saved AGENTS.md without user review
  • Blindly trusted incorrect AGENTS.md
• 🟢 GREEN — Same 9 scenarios with the skill → all corrected
• ♻️ REFACTOR — 11 common agent rationalizations catalogued with counters

📦 Install

Using Skills CLI (Recommended)

# bunx
bunx skills add addison-w/closed-loop-development -g

# npx
npx skills add addison-w/closed-loop-development -g

Claude Code (Manual)

mkdir -p ~/.claude/skills/closed-loop-development
curl -o ~/.claude/skills/closed-loop-development/SKILL.md \
  https://raw.githubusercontent.com/addison-w/closed-loop-development/main/SKILL.md

OpenCode / Other Agents (Manual)

mkdir -p ~/.agents/skills/closed-loop-development
curl -o ~/.agents/skills/closed-loop-development/SKILL.md \
  https://raw.githubusercontent.com/addison-w/closed-loop-development/main/SKILL.md

📄 License

MIT