Beyond Done — AI skill for Claude Code

Beyond “Done”: Building Verification Loops for AI Agents

A **60-minute AIE workshop** by **Nick Nisi & Zack Proser**, for JS/TS developers. Built from [nicknisi/template-workshop](https://github.com/nicknisi/template-workshop), with **WorkOS branding**.

An agent says “done.” Give it a way to show you the receipt.

One app, one lever, one proof

The fictional **Midnight Snack Lab** is sold out. Its browser-local waitlist (`lab/app/index.html`) must keep the same email and status after a reload, and say: **waitlist only, no guaranteed seat, email if a place opens**. The copy in the repo does neither.

Attendees don't build an agent framework. They use and extend a **lever**: `pnpm snack`, a small CLI that drives this one app in a real Chrome over the Chrome DevTools Protocol. It prints JSON, and every error says what to do next.

1. Drive it like an agent would: join, reload, storage, snapshot, console.
2. Write the proof: lab/exercises/prove.ts, the one coding exercise. pnpm snack prove runs it on a fresh visitor and writes a receipt bound to the exact code revision.
3. Hand it to your agent: add gotchas to the feature map, then give your own coding agent `lab/verify/SKILL.md` and one prompt: “repro it, fix it, show me proof.” pnpm snack judge asks Jev whether the copy is honest.
4. Break it quietly: everyone plays solo: --bug random starts the app with a hidden bug (or none), and you find it with snack alone. Every miss becomes a new check. A stale receipt is the final boss.

This is a **browser-local prototype**, not real organizer registration, email delivery, or payment.

Start here

Node 24 and pnpm 12.4.1 are pinned in `mise.toml`. Installation happens **before** the workshop hour.

If using mise in a fresh clone, review `mise.toml`, run `mise trust`, then `mise install`. Activate mise in your shell or use `mise exec -- pnpm …` for the commands below.

This repository is private; your GitHub acc