Scan to open the slides
workshop-cursor-deck.ai-builders.space
Learning to Code
When AI Does the
Heavy Lifting
Yuzheng Sun · April 1, 2026 · Room Willow A
Yuzheng Sun, PhD
Founder · Meta / Amazon / Statsig / Tencent
Build with AI: Mindset, Habits & Tools
for the Next 20 Years
Most people get 30% from AI. A few get 10×.
Sound familiar?
The difference isn't the tool. It's the method.
Bill Gates: AI is a paradigm shift on par with the GUI
"In my lifetime, I have seen two demonstrations of technology that struck me as revolutionary."
— Bill Gates, March 2023, after seeing GPT-4
Why this matters:
- Before GUI: only experts used computers
- After GUI: anyone could
- With AI: anyone can now harness compute
- What changes: what ordinary people can do
The electric motor: new technology ≠ swap the tool
When the electric motor was invented, factories simply replaced the steam engine with it — everything else stayed the same. Efficiency improved by 30%.
Decades later, someone realized: motors can be small. You don't need to design a factory around a central power source — you can design it around the production process itself.
The assembly line was born. Efficiency jumped tenfold.
Today's AI is the same story.
What most people do: keep the same workflow, just "ask AI" at one step.
— That's putting the electric motor where the steam engine was.
The real transformation: redesign how you work around what AI can actually do.
How to Use Cursor
ChatGPT is your advisor. Cursor is your engineering team.
Three fundamental limitations of chat-based AI
| Dimension | ChatGPT (chat mode) | Cursor (agentic mode) |
|---|---|---|
| Context | Every conversation starts from zero — you re-explain everything each time | Reads your entire project: files, rules, history — understands your setup |
| Persistence | Output lives in the chat window — disappears when you close it | Output goes into actual files — it accumulates, it's reusable next time |
| Self-correction | Only updates when you say it's wrong — silent on errors you don't notice | Has a real execution environment — runs code, sees errors, fixes itself |
Controlling the process vs. controlling the outcome
Traditional: Process Certainty
- Prompt → check → edit manually → repeat
- AI output capped by your review bandwidth
- You're the bottleneck
AI-Native: Outcome Certainty
- Define "what good looks like"
- Give AI rules + context
- AI runs, checks, iterates → you review
- Highest leverage: define the finish line
Let's open Cursor.
Feel the difference.
You don't need to
Already know how to code
Understand every AI concept
Learn everything today
You just need
One real task to work on
To tell Cursor your goal and context
To watch it run — then review
How to Debug
Red errors are clues, not disasters.
Debugging with Cursor: a 3-step mindset
Read the error — it's a clue
- Not a judgment — it's a message
- Tells you what broke and which line
- Read it before asking Cursor anything
Give Cursor the full picture
- Don't just paste the error
- "I expected X. Got this instead."
- Expected + actual = the right fix
Let Cursor run, check, and fix
- Let it execute, see output, and iterate
- Ask: "Why did this break?"
- Not just: "Does it work now?"
How to Write
Effective Prompts
A vague instruction can produce code with no errors — and still lose your data.
Download the practice data
We'll use this dataset to practice turning vague instructions into
precise, AI-ready specs with the 3C Framework.
What a vague instruction looks like in practice
You have a photo folder. Two files named sunset.jpg — one ocean, one mountain. You ask Cursor:
"Write a script to move all photos into one new folder, flatten the subdirectory structure."
THE TRAP
The code ran without errors. It did exactly what you asked.
But you lost data.
Vague requirements produce technically correct code that's a wrong product.
Why did this happen? The curse of knowledge
The problem isn't that AI is dumb. It's that you carry implicit assumptions that AI doesn't share.
AI: "Photos with the same filename are the same file."
AI: "If same filename = same file, then overwriting loses nothing."
The skill is not writing — it's excavating.
- Surface hidden assumptions
- Turn implicit decisions → explicit rules
- That's what Criteria is for
The 3C Framework: write specs, not requests
What's the #1 priority?
- Set the north star first
- Data safety: better to do nothing than lose a file
- This overrides all other decisions
What are the steps?
- Scan source directory (recursive)
- Identify duplicate files
- Execute file operations
- Generate verification report
Make implicit rules explicit
- Duplicate? → content hash, not filename
- Name conflict? → auto-rename, never overwrite
- Safety? → copy-only, originals stay
- Which files? → .jpg / .png only
Validation: trust the ledger, not your eyes
Data Consistency Check
- Add to end of every script
input = output + duplicates + skipped- If unequal → print RED WARNING
Visual Acceptance Report
- Generate
report.html - Left: files kept · Right: duplicates removed
- One human scan > 100 automated checks
THE MINDSET SHIFT
- "No errors" ≠ correct output
- Look for data integrity, not just no red text
- Define "done" before you run
Where you are now — what to do next
Only used ChatGPT
- Open Cursor on one weekly task
- Let it read your actual files
- Notice the context difference
No coding required. Just feel it.
Cursor but prompt one-by-one
- Write your standards in a file
- AI reads rules before acting
- Output quality jumps immediately
Already building workflows
- Add automated acceptance checks
- Second agent reviews / script validates
- "It works" → "It's repeatable"
This deck you just watched
is itself a demo
The process of making this deck practiced every principle we just covered.
How this deck was made
| Step | Human did | AI did |
|---|---|---|
| 1 · Brief | 1 message: workshop info + agenda + lesson content to include | Read original 747-line deck + lesson 4 materials + workshop screenshot |
| 2 · Content | Zero lines of HTML written | Output 900+ lines of HTML: 20 slides, full design system retained |
| 3 · Adaptation | Specified: English, debug section, 3C framework, Tencent in EN | Restructured narrative arc, added 3 new sections, repurposed all assets |
This is "outcome certainty" in a real workflow: human defines the goal and constraints, AI handles all execution — including the sentence you're reading right now.
Join the Superlinear Academy community
- AI Builder courses & projects
- Community feed + share your work
- Lifetime Q&A with instructors
Build with AI: Mindset, Habits
& Tools for the Next 20 Years
- Learn AI for 20 years, not the next 20 days
- Stop treating AI like a chatbot — design the context
- Build a meta-learning system that compounds over time
- Real projects, real code, real feedback from instructors
This deck is on GitHub
Every slide, every design decision, every prompt used to build this deck
is documented in the README — a live demo of the process.
The gap between using AI
and using AI well
isn't 30%. It's 10×.
Context · Components · Criteria.
The difference is method, not tools.
Yuzheng Sun · Superlinear Academy · Seattle (AI) Startup Summit · 2026