Everything I captured at AI Engineer World’s Fair 2026 — talk transcripts, wearable-recorder audio, a camera roll — turned into a keyboard-first spaced-repetition course, built almost entirely by an orchestrated fleet of Claude agents.
A personal learning platform that converts passive conference watching into the techniques that actually create durable knowledge for experts: retrieval practice on a spacing schedule, generation prompts, and build challenges. I started by researching the learning-science literature — retrieval practice, FSRS spaced repetition, the Feynman technique — and encoded those techniques as the product spec, so features fell out of the research instead of the other way around.
Then agents matched my camera roll to the conference schedule, wrote 66 lessons across two courses (the conference itself, plus a formal-methods course built from a folder of university PDFs), generated 274 retrieval cards, and typeset real MathML formulas. Everything compiles to one dependency-free HTML file with an FSRS-5 scheduler I implemented from the published weights.
Every one of these bugs ran green — exit code 0, pipeline “worked” — while producing wrong output. A passing pipeline tells you the machinery ran, not that the result is correct. The fix each time was to make an agent inspect the actual artifact.
The photo-to-session matcher assigned every photo to the wrong talk — macOS Photos returns capture times in your Mac’s current timezone, not the one you shot in. Caught only by making the agent read the slides visible in the photos.
ffmpeg exited 0, but no browser could decode the output: iPhone HDR is 10-bit (HLG/BT.2020) and libx264 happily kept it that way. Fix: tone-map to 8-bit BT.709.
A source PDF shared a name with a famous paper but was an unrelated one. A subagent caught it because it was instructed to flag source mismatches instead of paraphrasing blindly.
Agents write schema-validated lesson JSON, never HTML. Bad output fails validation instead of shipping broken markup.
A single deterministic generator owns every byte of HTML. Presentation is code, not something 20 agents each reinvent.
Append-only data changes so user review state survives a full regeneration of the course.
Save the big model for judgment calls; smaller models do extraction at scale. Fable 5 orchestrates, ~20 Sonnet subagents extract and write.
Playwright-driven headless Chrome as the agents’ verification loop — every change screenshot-checked and behavior-tested before it ships.
The working system: the review queue, conference photos matched to talks, clickable scholar bios, MathML lessons.
robbwinkle.zo.space/aiewf26-course/ → ResearchThe product & learning-science research that became the spec — retrieval practice, spacing, the expertise-reversal effect, what to steal from each tool.
Read the research → SpecWho it’s for, the product purpose, brand personality, anti-references, and the design principles — recall is the product.
Read the product spec → SpecThe design system: the Command-Marquee lineage, the three-job color rule, two type voices, and the one-light depth model.
Read the design system →