The Starting Point
It started with a single tweet — a video of Boris from Anthropic presenting Claude Code tips at a conference. The professor (Q) had an idea: "Can we turn this 28-minute talk into a full interactive course for high school students who've never coded?"
Q had two assets: the video and a personal wiki of 161 entries on AI tools, agent design, and engineering practices — accumulated over months of research. The challenge was turning raw knowledge into something a 15-year-old could follow.
The Build Timeline + Q's Actual Prompts
Each step shows what Q actually typed (the green bubbles) and what happened as a result. These are real, unedited prompts — this is what "directing AI" looks like.
Download & Transcribe the Video
Downloaded the video from X using yt-dlp. Extracted audio and transcribed with Whisper. 28 minutes of content became a 269-line transcript. Designed a 10-module course structure with Codex while Q slept.
claude_code_intro.txt + 10-module course planClaude Code Fundamentals
Built a self-contained HTML file with terminal simulator, quizzes, drag-drop exercises, and flashcards. Then Q woke up and expanded the scope:
Added GitHub, Web App Basics, and Deployment modules — growing the course from 10 to 13 modules. Codex reviewed 3 times, finding privacy inaccuracies, wrong install commands, accessibility gaps, and quiz XP farming bugs. All fixed.
claude-code-course.html (13 modules)Scan the Wiki & Plan Advanced Content
Scanned Q's 161-entry wiki. Categorized into 8 themes. Codex designed a 16-module advanced curriculum based on the "Five Levels of Mastery" framework from Q's own wiki entries.
161 entries scanned, 8 themes identifiedBuild Course 2 + Test with Personas
Q left for 2 hours. Built the Agentic Engineering Quest (10 modules, gamified, XP system). Codex tested with 3 simulated HS student personas: Maya (16, art-focused), Jason (17, gamer), Priya (15, debate skeptic). Fixed 12+ issues from persona feedback.
agentic-engineering-quest.html (10 modules)Connecting Beginner & Advanced
Q returned and wanted a clear progression. Created a landing page connecting both courses with a visual learning path.
index.html (landing page v1)Pre-Test & Post-Test
Embedded Q's existing assessment tool. Added pre-test and post-test links to the learning path. Later, Q decided to remove the mandatory gate — good products remove things too.
Catching Skipped Reviews
Q caught Claude skipping the Codex review step on a "small" change and held it accountable. This moment shows the human's role: enforcing process discipline even when AI rationalizes shortcuts.
Build Course 3: AI Skills Lab
Built 3 tracks from specific wiki entries: prompt frameworks (7 superpowers), AI as tutor (learning notes), and cross-model review (detective game). Each track produces an exportable artifact.
skills-lab.html (3 tracks, 11 modules)Scanning All 161 Wiki Entries
Q noticed only 40 of 161 entries had been scanned and pushed for completeness. Three parallel agents scanned the remaining 121 entries. Found 38 total that were accessible to HS students.
38 HS-accessible entries from 161 total6 Courses vs. 1 Combined Course
Claude proposed 6 mini-courses. Codex pushed back: "scope creep." Claude countered: "pure study skills feels preachy — add a creative track as a reward." Q forced them to negotiate until both agreed: 1 combined course with 3 tracks (Study Better + Think Critically + Create Cool Stuff).
6 courses → 1 combined course (consensus)AI Life Skills
13 modules across 3 tracks. Study Better (Socratic tutor, cognitive fatigue, deep prompts), Think Critically (hallucinations, benchmarks, safety, independence), Create Cool Stuff (game design, video planning, story writing).
ai-life-skills.html (3 tracks, 13 modules)4 Major Iterations
Started as simple card list → added orbital hero → Q rejected dark theme → generated impressionist painting → frosted glass nodes → fixed click bugs → added recommended learning path. The landing page went through more iterations than any other file.
index.html (4 major iterations)Cat Mascot & Icons
Q shared a photo of their tuxedo cat. AI generated a logo: the cat wearing round glasses, sitting next to a laptop with a brain icon. Also generated 6 themed course icons — all with matching glow effects.
logo.png + 6 icon filesGoing Live
Deployed to GitHub Pages. Debugged caching issues. Q decided to remove the pre-test gate — simplifying the user experience.
Live at weihaoqu.github.io/learnai-courseRecording the Process
This page. Q wanted full transparency — not just the result, but the messy, iterative, real process of building with AI.
case-study.html (this page)Unifying the Academy
Codex audited all 6 files and found two different dark themes and a completely different light theme in Course 1. Added a consistent nav bar across all pages, standardized colors, and made the case study a 7th orbital node. Then Q asked for watercolor-style icons to match the impressionist painting — regenerated all 7 icons in that style, and increased node sizes for readability. The final round of polish: taste, cohesion, details.
Consistent nav + watercolor icons + enlarged nodesFilling Gaps and Fixing Bugs
The biggest expansion session. Q provided 12 YouTube links; Claude analyzed each video's transcript, and Codex reviewed placement decisions. 12 videos were embedded across all pages: an "AI for beginners" intro on the landing page, Claude Code and visual workflow demos in Course 1, an optional AI agent builder in Course 2, the "vibe learning" warning and LLM explainer in Course 4, and a dedicated "How AI Actually Works" background page with 4 progressive videos (including a Chinese Stanford course summary with English subtitle instructions).
A Chinese prompt engineering tutorial was adapted into a downloadable cheat sheet for Course 3. Codex then ran a full QA audit simulating two student paths and found a critical infinite recursion bug in Course 3's navigation — switchTrack() and goMod() called each other endlessly, crashing the browser on any button click. Fixed by extracting a separate showMod() function. Also fixed a missing JS function reference, an HTML injection vulnerability, and a placeholder link.
Adding Depth to the Experience
Q found a Three.js dotted wave React component and wanted it on the course pages. Codex rejected Three.js (600KB, WebGL startup lag on Chromebooks) and proposed a dependency-free 2D canvas version using fake 3D perspective projection — same visual, zero dependencies, ~3KB. The first version only covered the top band; after iterating on camera angle, scene depth, and grid density, the wave expanded to fill the full viewport as a subtle fixed wallpaper. Q and Codex tested three color/opacity options; Option C (cyan-light at 16% opacity, smaller dots) won for best contrast without competing with content.
wave-bg.js: 3D dotted wave, zero dependencies, shared across 5 pagesKey Decisions We Made
Building is choosing. Here are the real tradeoffs we debated:
React App vs. Self-Contained HTML?
Why: The academy's biggest advantage is accessibility. A teacher can share a GitHub Pages link and every student can use it immediately. Adding React would break that for one visual effect.
6 Mini-Courses vs. 1 Combined Course?
Why: Codex argued that 6 courses turns a clear learning path into a content maze. Students need depth, not breadth. One strong course beats six thin ones.
Engineers vs. High School Students?
Why: "Harness engineering" means nothing to a 15-year-old. "Set your agent up to succeed" means everything. Same concept, different framing.
Dark Theme vs. Impressionist Painting?
Why: Q's taste. The professor wanted something that said "creativity and learning" not "hacker terminal." The painting sets a completely different emotional tone.
What the Human Did vs. What the AI Did
This is the most important section. Building with AI doesn't mean the AI did everything.
👤 The Human (Q)
- Set the vision: "I want non-CS HS students to learn agentic engineering"
- Chose the audience and adapted scope
- Pushed back on design decisions (dark → painting, pixel → round particles)
- Judged taste: "this is ugly", "the preset is too small", "I like the previous version better"
- Enforced quality gates (Codex review rule)
- Provided source material (161 wiki entries, cat photo)
- Made final ship/no-ship decisions
- Caught when Codex review was skipped
- Defined the pedagogical philosophy
- Curated 12 video links and decided which to include
- Recruited testers and reported bugs
🤖 The AI (Claude + Codex)
- Transcribed video (Whisper)
- Designed course structures and module plans
- Generated and revised all HTML/CSS/JS code
- Reviewed every build for bugs and content accuracy
- Simulated 3 student personas for testing
- Scanned and categorized 161 wiki entries
- Debated scope (6 courses → 1) and reached consensus
- Generated logo and course icons from a cat photo
- Generated impressionist painting background
- Fixed dozens of bugs and issues across multiple review rounds
- Analyzed 12 YouTube video transcripts and recommended placements
- Ran full QA audit simulating student walkthroughs — caught critical recursion bug
- Searched the web for content gaps and recommended new videos
Lessons Learned
Try This Yourself
You don't need to be a professor or have 161 wiki entries. Here's a simplified version of what we did:
- Pick a topic you know well — a hobby, a school subject, a skill
- Find one source — a YouTube video, a blog post, your own notes
- Ask AI to design a 3-module course — specify your audience (who's learning?)
- Build it — ask AI to create an interactive HTML page with quizzes
- Review it — read every word. Fix what's wrong. Ask a friend to try it.
- Ship it — share the HTML file with your teacher, or put it on GitHub Pages (free)
A small first version can take 2-3 hours. A polished version takes more iteration. But the result is something real that helps real people learn.