Peter Yang built a self-improving AI life coach inside Codex using five structured markdown files in under 13 minutes. The system runs on Claude via Anthropic's Fable and uses four core files: skill.md defines behavior, plan.md stores goals and life context, learnings.md accumulates patterns over time, and eval.md sets quality checks the AI runs before responding.
The architecture is worth understanding beyond the demo. The learnings.md file is the mechanism that makes this self-improving rather than static. Every interaction can update that file, meaning the advisor gets more calibrated to you specifically over time. The eval.md layer forces the AI to audit its own output before delivery, which is a concrete implementation of self-critique that most prompt tutorials skip entirely.
The written tutorial is linked in the description and goes deeper on each file's structure. If you are building any kind of personal productivity layer on top of Claude or similar models, the file-based memory pattern here is directly reusable. The specific combination of a context file, a memory file, and an eval gate is the part worth studying.
[WATCH ON YOUTUBE →]