Learn It by Building It: the Agent Skills Course Site

Background

The course site teaches Matt Pocock's agent skills — 25 of them, five chapters. It is also its own case study: I didn't learn the skills first and build afterwards, I built with them while learning them. The repo was set up with /setup-matt-pocock-skills on Sep 18, the first full pipeline run finished that same day, and the build wrapped on Sep 20. Three days. Every workflow the course teaches has one real run in the repo's issue history, and you can click through all of them.

My role

One person, two hats. As a learner I kept test scores, weak-cluster diagnoses and prescriptions in a daily log — ten entries under learning-records/. As an operator I ran the pipeline that took the site from idea to launched product. Both hats share one evidence trail: GitHub issues #1–#24, one commit per issue, searchable in git log as feat: 工单 #N.

Technical decisions

The main pipeline — grill → spec → tickets → implement → code-review, with implement driving /tdd one red-green slice at a time — ran end to end three times: the course home (#1 spec → #2/#3 tickets), the dictionary (#4 → #5/#6), and the public-site refactor (#12 → #13–#24, twelve tickets with blocked-by dependency edges; nine landed the same day). Every ticket ran in a fresh session, context cleared when done.

Besides the three specs, #7 was a different kind of ticket. Turning a private study workspace into a public course site: audience, information architecture, register, vehicle — none of it decidable, fog too thick for one session. So /wayfinder drew a map. Two research tickets (#8, #9) went to background subagents in parallel; two decision tickets (#10, #11) I answered question by question, and for the first time I pushed back on the AI's proposal — navigation experience and written register — which reshaped the map. Before deciding, /prototype produced three candidate homepages, and the "answer experiment" became the entire content of #10. The day the map cleared, /to-spec wrote #12 and /to-tickets split it into twelve tickets.

One decision got overturned along the way: #7 settled on pure static + GitHub Pages, and when #12 was written it became Next.js + shadcn/ui + Vercel, revision archived. The map governs direction, not loyalty.

The stack is unremarkable on purpose: Next.js + TypeScript + MDX, statically exported, frontmatter validated at build time. What mattered was the testing seam — "done" was judged by a machine against the rendered static output, not by feel.

Outcome

Launch landed on day two, followed by three failures in a row: the Vercel import lacked a Root Directory so the build died; statically exported RSC prefetches 404'd, so every Link dropped prefetch; the resume-learning button mismatched hydration, fixed with a two-stage default. The one-off Vercel token issued during debugging got a /wizard-generated four-phase script (locate → verify → y/N-gated delete → confirm) and was revoked once its life ended. Then /improve-codebase-architecture ran the first architecture review; six improvements landed, including a real latent bug: the MDX pipeline never had GFM enabled, so tables in the first four lessons rendered as vertical bars.

The neat part is the loop. The wayfinder lesson went live while the same map was being drawn for this very site, and post-launch bugs and architecture debt became practice material for the upkeep chapter — the courseware grew out of the site's real decisions. After day three the loop kept running: spaced retests kept producing data (records 0007–0010), the weak cluster was re-diagnosed from "discipline no-go zone" to "old intuition overriding the new rule", and quiz generation started adding deliberate variations.

The full retrospective lives in the repo, three layers deep, all clickable: docs/journey.md (timeline plus a skill-by-skill evidence table), issues #1–#24, and learning-records/. Claims about this workflow should be auditable, not aspirational — that trail is the audit.