Every lesson is an engineering ticket on Project Forge. Approach each as real work: understand what's being asked, build it in layers with clear contracts, validate the untrusted boundary, make security fail safe, type-check the contracts, test the logic, exercise the endpoints, and document your reasoning. The goal isn't memorizing a framework — it's the architecture, security, and judgment to ship a backend a team can put on call.
- The boundary is untrusted. Validate every request before any logic runs; never trust
req.body. - Layers with clear contracts. Controller (HTTP) / service (logic, no HTTP/DB) / repository (data behind an interface).
- Make illegal states unrepresentable (from Module 04): typed domain, literal unions, domain errors mapped to status codes in one place.
- Fail safe, not open; least privilege. Security defaults to deny; an unverified token is no identity.
- Measure first, design for failure (from Module 03): health/readiness, retries/backoff, dead-lettering, graceful shutdown.
- Read the ticket and the deep dive.
- Do the lab. Build it, predict the type-check/test/status-code result, then verify.
- Investigate — push from "it returns 200" to "I can show the bad input rejected before the service, the token failing safe, the 403-vs-401, the idempotent retry acting once, the readiness flipping when a dependency dies."
- Run the AI exercise — draft → verify → log, deliberately.
- Submit the assignment and update your notebook.
- Check the solution to validate your reasoning — after you've done the work.
Track progress in dashboard.html.
- What you built and why this shape — which layer holds the logic, how the boundary is validated, how security fails safe, what's idempotent.
- Evidence: the type-check result and confirmed contract errors; Node test output (logic + auth via real crypto); endpoint request/response transcripts (status codes, idempotency, 401/403, 429, readiness).
- The fix at the cause — no
any, no silenced errors, no trusting input. - AI-usage log: draft → verify → log.
- Clean commits (Module 02 habits).
AI drafts fast and confidently, and is sometimes wrong in ways that are dangerous on a backend — trusting input, fast-hashing passwords, trusting a token before verifying, allow-by-default authorization. You have three mechanical checks: the type-checker, Node tests (auth with real crypto), and real requests. resources/ai-workflow-guide.md maps the failure modes.
A contract that doesn't type-check isn't done; a security claim without a test isn't proven; an endpoint behavior without a real request isn't verified. The compiler, the tests, and the requests are the arbiters — not confidence. Fail safe, not open, and prove it.
Against ASSESSMENT_RUBRIC.md — on architecture, security, API platform quality, data, background processing, and operations, with evidence. An endpoint that returns 200 but trusts its input, fast-hashes passwords, or authorizes by default scores poorly regardless of the happy path.