Vibe
Coding
Learn the latest AI-based tools for software development — by building. Ten Thursdays, ten builds: from a single prompt to a shipped app.
Thursdays 10:00–11:20 · Savery Hall 220 · Steve Seitz
about
The goal: build your own Claude Code
By the end of the quarter, you won't just use AI coding tools — you'll have built one: your own coding agent, with a harness, tools, and guardrails you understand line by line.
Each 80-minute session splits in two: a lecture and live demo of the week's core technique, then roughly 40 minutes of in-class building with the instructor and TAs. Week by week you assemble the pieces — the agentic loop, specification and steering, MCP tools and governance, multi-agent orchestration, evals — until they compound into an agent that's yours.
schedule
Ten Thursdays, ten builds
The sequence starts with the fastest possible win — a prompt becomes a working app in session one — then climbs through prompting method, the agentic loop, agents, and closes on evals and shipping.
- L01Lovable / Bolt / v0
Prompt to App
Course overview, a short history of AI coding, and your first end-to-end app — prompted, built, and shipped in session one.
- L02VS Code + Claude Code
Coding on a Budget
Model choice as a hiring decision: capability, task fit, and total cost. The prompt-run-refine loop, practiced by hand.
- L03Python + Claude on Bedrock
The Agent Harness
The agentic loop, named precisely — observe, decide, act, stop. You build a hundred-line coding agent yourself.
- L04Claude Code
Steering a Coding Agent
Specification, planning, interruption, instruction files, and skills — the surfaces that steer a delegated agent.
- L05Claude Code + MCP
MCP & Agent Governance
Connect third-party tools through MCP, then govern them: permissions, allowlists, sandboxing, audit trails.
- L06VS Code, agent mode unlocked
Working in Code You Did Not Write
Most industry AI work happens in someone else's repo. Reviewing, testing, and security-auditing AI-generated code.
- L07CrewAI
Multi-Agent Orchestration
Planner, workers, critic. Supervisor, fan-out, pipeline, debate — what multi-agent buys and what it costs.
- L08Ollama
Local Models
Run models on your own machine: privacy, offline, zero marginal cost. Structured outputs and local-first routing.
- L09Promptfoo
Evals
Design evals that catch real failure modes. Compare models with your own harness and audit a quarter of your own work.
- L10GitHub Pages + Actions
Deploy Behind CI
Ship an earlier project: deploy targets, smoke tests per pull request, and recurring automated workflows.
grading
Fair, repeatable, auto-graded
Projects are graded against a standard rubric by analyzing the code and the generation trace — two students doing similar work score about the same.
Weekly builds, 10 points each, plus 1 bonus point for the most polished or creative submission — demoed at the start of the next lecture.
Weekly auto-graded quizzes on the core concepts and the readings.
Includes the short feedback survey after each lecture — your input literally reshapes the course.
bonus: out-of-class AI artifacts earn extra credit, with awards at the end of the quarter and short demos during lectures.
staff
The instructional team
Steve Seitz
Professor · Instructor
Professor in UW CSE working on computer vision, graphics, and generative AI — and a VP Fellow at Google, where he has led teams building products like Beam and Flow. He normally teaches the computer graphics course.
seitz@cs.washington.eduTeaching assistants
- VAVinamra Agarwal
- ECElla Cao
- PBPrabhgun Basi
- ASArian Shamaei
- AKAditya Kumar
Every TA here has serious vibe-coding mileage — bring them your questions.