From idea → deployed production app, by directing coding agents — reading ~10% of the code, strategically. Not a CS degree. The director's chair, with just enough fundamentals to supervise well.
Every module advances one real build: THG Radar — an internal tool tracking client integration go-lives. Cloudflare Worker + Hono API + D1 database + plain HTML UI, locked behind Zero Trust, deployed to production. M1 scaffolds it; M8 ships it; M9 gives it an AI feature; M10 fences its spend. The capstone: you ship one more feature, alone, via an agent — correctly.
Right now an idea becomes software by convincing an engineering team it's worth their sprint. After this track, an idea becomes software because you sat down with a coding agent for an evening — and what came out is deployed, behind auth, tested, and won't quietly burn money. You'll read maybe a tenth of the code. You'll know exactly which tenth, and why that's enough.
Full loop solo: spec → plan → diff review → tests → deploy → verify. Pass: it's live, and you can answer the 3 questions about every diff you accepted.
The whole track as a clickable tree — AI Engineering → 4 branches, fundamentals expanding to 5. Click any node: what it means, why it matters, a hands-on drill. Keyboard accessible.
Same feature, vague vs sharp spec → visibly different plausible agent outputs, divergences called out. Output quality is set before the agent types.
Real-ish diffs, some with a planted bug or leaked secret. Find it, get explained feedback. This is what makes "not reading all the code" safe.