🔁 The Self-Improving Loop

Teach your coding agent to stop making the same mistake twice. For AI builders, built from a live app's real process files. Pick how deep you want to go.

⚡ Just the concept 5 minutes · one page

Why agents forget, the loop in one picture, the two ideas that make it work, and one prompt to start tomorrow.

Start the 5-minute version → 🔬 The deep dive 8 lessons · about an hour

Every station in detail, with real stories, practice and a short quiz per lesson, ending with a starter kit you install in your own project.

Start lesson 1 →

What you will have at the end of the deep dive

A loop running in your own project: your agent writes a retro after each release, lessons land in a ledger with deadlines, the ones that matter become numbered rules every session loads, and the most important ones become checks that block the mistake before it ships. Works with Claude Code first, and with Codex or any agent that reads AGENTS.md.

The deep dive

  1. 01
    Why your agent keeps repeating mistakesFresh sessions, advice vs enforcement, the practices the loop borrows, the loop in one picture.
  2. 02
    The retro: let the agent write the post-mortemBlameless; trigger, root cause and detection gap; gems, friction and rules cited.
  3. 03
    The ledger: one row per lessonStatus, source, notes; a working queue that prunes itself.
  4. 04
    Triage: make "later" expireAdopt, reject, trial(date), deferred(date), and the deadline that forces the decision.
  5. 05
    The playbook and promotionOne agent-agnostic source of truth; thin CLAUDE.md and AGENTS.md; numbered rules.
  6. 06
    Make it impossible: the enforcement ladderProse, checklist, test, probe, hook, server rule; failing closed.
  7. 07
    Keep the loop honestRule citations, demotion, trials, a second model auditing the retros.
  8. 08
    Install the starter kitThe files, one prompt for your agent, and your first-week routine.

📘 Reference

Glossary: the course's words, one definition each.

Case study: HallMingle, a live multiplayer meetup app built with coding agents, whose learning loop has run since July 2026. Stories are real and sanitized. Sources checked 2026-10-05.