Session 5 · Friday, August 28, 2026

AI for Building Apps

Han Zhang and Bing Tian

Today we pull these five days into one small build, and we look at what comes next. When a section has something to do, it appears in the panel on the right.

Opening

Today

  • The goal today is to pull these five days into one small build, and to look at what comes next.
  • Here is the plan for the day:
Turn your analysis into a website or app
Build one of your own (100 minutes)
Break
Keeping private data on your own machine
Show your project
What to learn next

Turn your analysis into a website or app · do it on the right

What building means here

  • Building means turning your analysis into a website or an app that other people can use:
    • a webpage that shows your results and lets people click through them
    • one command that reruns your whole analysis, from raw data to finished tables
    • a small app that does a repeated chore for you
  • The division of labor is the same as Monday. You specify and verify, and the agent writes the code.

How a web page works

  • Every webpage, from a personal site to a results page, has three layers:
    • what the reader sees, which is the structure, the style, and the behavior
    • what changes on the page, which is the inputs and filters
    • where the numbers come from, which is your analysis output
  • You will let the agent write all of these layers.
  • You need to know the layers so you can check the agent's work. Open the page, click through it, and verify the numbers against your output.
Live demo

From an analysis to a page, in minutes

  • Tuesday's analysis becomes a small interactive page, live.
  • The prompt names the output file that feeds the page and the one control the reader gets.
  • The page opens in a browser within minutes, and I have read none of its code.
  • The check at the end is the number on the page against the number in output/.

Keeping private data on your own machine · do it on the right

When your data cannot leave your computer

  • Some data cannot be sent to a company's servers, such as human subjects, restricted records, and anything under an IRB or a data agreement.
  • The default tools send your files to Anthropic or OpenAI, so for that data you cannot use them.
  • The fix is to run the tool on an open-source model on your own machine, so nothing leaves it.

How to keep it local

  • Ollama is a free program that runs open-source models such as Llama or Qwen on your laptop.
  • You can run Codex with its open-source flag, or point Claude Code at the local model with three settings: ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN, and ANTHROPIC_MODEL.
  • You set the sandbox to your project folder and turn the network off.
  • Local models are slower and weaker than Claude or GPT, so you use them only when the data requires it.

Show your project

Show your project

  • Take a few minutes each to walk us through your whole five-day project, including today's build.
  • Show the question, and show what you have: a result, a figure, a website, or a working start.
  • Walk the main steps and say who did each: what you decided and did, and what the agent did.
  • Tell us one thing the agent got wrong, and how you caught it.
  • Leave us one lesson for someone starting next week.

What to learn next

Your five days, as a checklist

  • a working environment (Monday)
  • a research repository with a draft history (Monday onward)
  • a directed, verified analysis (Tuesday)
  • a project AGENTS.md that survives a restart (Wednesday)
  • at least one Skill (Wednesday)
  • a literature review or draft with verified citations (Thursday)
  • a website or app of your own (Friday)

What the full course adds

  • How these models actually work, and why they fail the way they do.
  • Designing your own agents with MCP, Hooks, and subagents. The next cards preview two pieces of this.
  • Building and deploying real tools on top of your research.

Who splits the work?

  • Today's capstone fit in one conversation, and your next real project will not.
  • The design question that follows is decomposition: when a task is too big for one session, who breaks it into pieces, you or the agent?
  • Claude Code covers the full range:
    • You decompose fully. A shell loop calls claude -p once per item on a list you wrote, when you already know every piece.
    • You plan, the agent executes. Plan mode makes it present a plan you approve before it touches a file, when you know the goal and the steps need thought.
    • You frame, the agent splits. "Do A, B, and C in parallel with separate subagents" works when the pieces are independent and summaries are enough, as Thursday's paper fan-out showed.
    • The agent decomposes fully. A workflow has the agent write an orchestration script that runs dozens of subagents, for a repository-wide audit or a hundred-item migration.

Before you parallelize

  • Four rules of thumb from Anthropic's own multi-agent engineering:
    • Independence. Only subtasks that do not depend on each other's output can run in parallel. A chain of dependent steps gains nothing.
    • Cost. A multi-agent run measures at about 15 times the tokens of a single chat, so the task has to be worth it.
    • Focus. Three agents with sharp, separate briefs beat five with scattered ones.
    • Fresh eyes. A reviewer agent with clean context catches what the author's context has learned to overlook, for the same reason journals use outside referees.

After the five days

  • Keep one repository per project, and let it grow the way this one did.
  • Reuse your AGENTS.md and Skills on the next project from day one.
  • When you hit a task too big for one session, that is when subagents and workflows earn their keep.

How good, and why a second one?

  • How good, honestly. Small local models in the 7B class drop 15 to 40 points on agentic benchmarks and hallucinate tool calls. Frontier open-weights models such as Kimi K2.6 and Qwen 3.6 sit within a few points of the closed frontier.
  • A second use, cross-model review. BK Lee's academic setup wires /codex and /gemini commands into Claude Code, so a second model family reviews the first's work: "each catches what the others miss."
  • The statistical logic. Error cancellation assumes independent errors, and models from one family make correlated ones. A reviewer from a different family is closer to an independent draw.

Keep going

  • Keep the repository alive: one task per week through the agent, and the AGENTS.md grows with the project.
  • Free self-paced material if you want more:
    • Anthropic Academy: short courses on Claude Code, from beginner up.
    • DeepLearning.AI: "Claude Code: A Highly Agentic Coding Assistant."
  • The habit that matters more than any tool: an unverified result is not a result.