A flexible template for semi-autonomous AI coding. Clone it, shape it to your needs, and start building - with verification and guardrails built in. No packages to install, no services to configure.
Low mental overhead for you. Commands and verification keep work on track without constant intervention.
Guided Q&A extracts clear requirements. Every decision is captured, every edge case surfaced.
Generator creates tasks with testable acceptance criteria. Every task is verifiable—no ambiguity.
AI agents work task-by-task. Automatic verification after each one. Human checkpoints between phases.
Built-in verification, security, and recovery. No more crossing your fingers.
Sub-agents verify each task against acceptance criteria. Code doesn't pass until tests pass.
Verifies tests exist, were written first, and have meaningful assertions. Git history proves compliance.
Dependency audits, secrets detection, and static analysis at every checkpoint. Blocks on critical issues.
Agents escalate to humans after repeated failures instead of spinning forever. No more burning tokens.
Analyze failures, rollback phases, retry tasks with fresh context. Fix problems without starting over.
One branch per phase, one commit per task. Human review before push. Clean, traceable history.
Checks that requirements flow through document chain. Nothing lost between product spec and implementation.
After checkpoints pass, automatically advance to the next phase. Hands-off execution with a 15-second interrupt window.
Save project-specific patterns and gotchas to LEARNINGS.md. Future sessions load this context automatically.
When you say "v2" or "later" during Q&A, the toolkit captures it with context. Nothing gets lost in the shuffle.
Works with Claude Code and OpenAI Codex CLI. Auto-detects your environment and adapts accordingly.
I started using CLI tools (Claude Code, Codex CLI) after vibe coding tools like Replit and Lovable left me wanting more.
Using these CLI tools for some time, I, like many people, wanted to improve my workflow. AI coding agents are powerful, but unreliable. The CLI tools help govern the agents, but with that freedom also comes the risk of adding overhead and complexity.
With most of the workflows that I found online and in AI coding communities, they were both complex (87 agents! Ralph! Run 10 Claude's at once!), and they were relatively opinionated. I found these workflows to be too restrictive, particularly in the dynamic, fast-moving AI coding space.
This workflow has a few goals:
That's why it's a Git repo that you can clone, fork, inspect, and modify as you see fit.
If it's working well, you'll be able to understand & customize your workflow; you'll spend more time planning than pasting; and you'll spend more time reviewing completed work, than inspecting half-completed prompts.
We used this toolkit to build a calculator app from scratch. Browse the generated specs, execution plan, and final working code.
No npm install. No configuration. No accounts. Just clone the repo and start using the slash commands in Claude Code.
Open source template. Clone it, customize it, make it yours. No installation required.