Building with an AI coding agent
Danny Weagley's set-up guide for build-round teams. It covers what Claude Code and Codex are, where your team's files live, the steps to get a shared repo running, and seven prompts to copy. Danny walks through it at the October 23 build session; this page is the version you can keep open.
Prepared by Danny Weagley.
What these tools are
You have probably used Claude or ChatGPT in a chat window: you ask, it answers. Claude Code and Codex are different. They work inside a folder on your computer. They can create files, edit them, run the program, see the error, and fix it, all on their own. You describe what you want; the agent does the typing.
Your job (the team)
- Decide what to build and for whom
- Approve the plan before it builds
- Try every change yourself
- Check that everything is right
- Save your work and keep the session log
The agent's job
- Propose a plan when you ask
- Write and edit the code files
- Run the code and read the errors
- Explain what it did in plain English
- Write the save message for GitHub
You do not need to read code to use these tools. You do need to test what they build and be able to explain it; judges will ask.
The build loop
Every working session follows the same six steps. Small steps are the secret: one feature at a time, tested before the next.
- 1Describe
Say what you want in plain words. Who uses it, what it does.
- 2Plan
Ask for a plan first. Read it. Say OK or change it.
- 3Build
Let it build one step of the plan, not all of it.
- 4Test
Open it and try it yourself. Try a number you can check by hand.
- 5Fix
If it breaks, paste the exact error. One change at a time.
- 6Save
Write the session log, then commit to GitHub.
Then go back to step 1 for the next feature. Think and plan in chat; build in the agent.
Terms you'll hear
- Folder / project
- The one folder on your computer that holds your app. The agent works only in it.
- File
- One piece of the app, such as index.html (the page) or app.js (the logic).
- GitHub
- A website that stores your team's code online, keeps every past version, and lets teammates share it.
- Repository (repo)
- Your project's home on GitHub, one per team. Each teammate keeps a copy (a "clone") on their own computer, pushes their changes up to the repo, and pulls teammates' changes down.
- Commit
- A saved snapshot of your work, with a short note on what changed. You can always go back to one.
- Pull / push
- Pull = get your teammates' latest changes. Push = send your commits up to GitHub.
- Deploy
- Put the app on the internet with a link anyone can open (Vercel does this for free).
- API key
- A password that lets your app use a data service (FRED, CoinGecko). Treat it like a password.
- Terminal
- A text window for typing commands. Optional: the desktop Claude/ChatGPT apps let you skip it.
More prompts in the prompt library. Start with How the workflow works, Put your work on GitHub, Ship it with Vercel, and the Make your test data track.
Where your team's things live
| What | Where | Rule |
|---|---|---|
| Code (the app itself) | A GitHub repo, one per team | Pull before you start; commit and push when you stop. Never keep the code folder inside OneDrive: syncing breaks it. |
| Session logs | A logs/ folder in the repo | One short file per person per session: Did / Decided / Open / Next. The agent reads the latest ones when you start. |
| Slides and video | Your team's OneDrive or Google Drive folder | For the files that are not code. |
| Thinking and planning | Claude chat (UT AI Hub) or ChatGPT | Brainstorm, compare ideas, draft the plan. |
| Building | Claude Code or Codex | Pick one tool as a team so your setup matches. |
| Team instructions for the agent | A file called AGENTS.md in the repo | What the app is, how to run it, your rules. Every teammate's agent reads it first. |
| Live app | Vercel (free) | Connect it to GitHub once; every push goes live. |
Two people, one file, same time = trouble. Split the work by feature, and say in your group chat who is working on what.
How the pieces connect
Follow the numbers through one work session. The outlined box is your own computer; everything outside it is shared with your team.
- Plan in chat. Talk the idea through in Claude chat. Leave with a short plan.
- Describe it to the agent. Give Claude Code or Codex one step of the plan.
- The agent builds. It writes and edits files in your project folder and runs them.
- You test it. Open the app, try it, and check one number by hand. Not right? Back to 2.
- Push to GitHub. Save a commit and send it up, so the team's shared copy has it. Pull first, so you start from your teammates' latest work.
- Teammates pull. They get your changes on their laptops and push theirs the same way.
- Vercel publishes. Every push to GitHub updates the live link automatically.
- Wrap up. The agent writes your session log (Did / Decided / Open / Next) into logs/, then commits and pushes it.
Pull before you start, push when you stop. Then nobody overwrites anybody.
Code and logs in GitHub; slides and video in OneDrive or Drive. Never put the code folder in OneDrive.
Keys stay on your laptop. The .env file never goes to GitHub or into a chat.
Set up step by step
Do this once, as a team. The team lead does the steps marked Lead; everyone does the steps marked Everyone. Plan on 45 minutes. Do the parts in order: A, B, C, D.
Your AI tool: account, plan, app
Pick one tool as a team: Claude Code (from Anthropic, maker of Claude) or Codex (from OpenAI, maker of ChatGPT). The steps below are for Claude; Codex works the same way.
Everyone- Make an account. Go to claude.ai → Sign up, using your UT email. (Codex: chatgpt.com.) The free UT AI Hub is for chat only; the build tool needs a paid personal plan.
- Pick a plan. In claude.ai: your name (bottom left) → Settings → Billing, or the Upgrade button. Organizers suggest the Max plan for the two months of the build round. Build-round students have the scholarship instructions by email.
- Protect your card. Plans renew every month until you cancel. Put a reminder in your calendar to cancel during month two (you keep access to the end of the month). Leave any extra usage option off, so you are never charged past the plan price.
- Download the app. Go to claude.ai/download, pick Mac or Windows, install it, and sign in. (Codex: install the Codex app from OpenAI and sign in with your ChatGPT account.)
- Find the Code tab. Open the Claude app and switch from Chat to Code. Chat is for talking; Code is the agent that works in a folder. It will ask you to choose a project folder: you will make that folder in part B, step 7.
- Allow it to work. The first time, the app asks permission to read and edit files in that folder. Say yes for your project folder only.
GitHub: your team's shared code
- Sign up at github.com with your UT email. Optional: apply for the free GitHub Student Developer Pack (education.github.com).
- Install GitHub Desktop (desktop.github.com). It gives you buttons for pull and push, so you never need the terminal.
- Make one folder for code that is not inside OneDrive: C:\code on Windows, or a code folder in your home folder on a Mac.
- On github.com: + → New repository. Name it after your project, choose Private, tick Add a README, and under ".gitignore" pick Node (it already blocks .env).
- In the repo: Settings → Collaborators → Add people, and add each teammate's GitHub username.
- Accept the invitation (check your email).
- In GitHub Desktop: File → Clone repository, pick your team's repo, and set the local path to the code folder from step 3.
- In the Claude app's Code tab (or Codex), choose this cloned folder as your project folder. It is the "project folder" in How the pieces connect.
OneDrive: slides and video
- Sign in at onedrive.com with your UT account. New → Folder, named after the team.
- Share → enter each teammate's UT email → set to Can edit → Send.
- Inside it, create folders for slides and video. (Session logs live in the repo, not here.)
- Open the link from the email. Click Add shortcut to My files, so the folder shows up in your own OneDrive.
- Keep documents here. Keep code out of here.
AGENTS.md: instructions for the agent
Agents need context about your project: what it is, how to run it, and your rules. They get it by reading a Markdown file, a plain text file with simple formatting whose name ends in .md. Claude Code looks for CLAUDE.md; Codex (ChatGPT) looks for AGENTS.md.
Put it at the top level of the repo, next to the README, not in a subfolder. Every teammate gets it when they pull, so every agent starts with the same background.
Claude Code reads CLAUDE.md; its one line pulls in AGENTS.md, so both tools follow the same instructions.
- Lead: open the agent in the cloned folder and run prompt 1 from Seven prompts to copy. It writes AGENTS.md with your session routine and makes the logs/ folder. Also ask: "Create CLAUDE.md containing only the line @AGENTS.md."
- Read AGENTS.md. Fix anything wrong. Commit and push.
- Everyone: pull. From now on, type start when you sit down and wrap up when you stop. Update AGENTS.md whenever the team decides something the agent should know.
Before you leave the first session
Each person checks every box. If one fails, a mentor can help before you go.
Seven prompts to copy
Prompt 1 is run once, by the team lead, during setup. Prompts 2 and 7 open and close every session; once your AGENTS.md has the routine, typing start or wrap up is enough.
1 · Start the project
Lead only · onceCreate AGENTS.md for this project: [one paragraph about the app]. Include how to run it; our rules (fake data only, API keys only in a .env file that is never committed); and a session routine. On "start": pull, read AGENTS.md and the three newest files in logs/, then propose a plan. On "wrap up": write logs/[date]_[name].md with Did / Decided / Open / Next, update AGENTS.md if needed, then commit and push. Create logs/ with a short README.md explaining this.
2 · Start a session
Pull the latest changes. Read AGENTS.md and the three most recent files in logs/. Tell me where the project stands, what is open, and propose a plan for today.
3 · Plan first
Before writing any code, propose a plan. Start with the smallest version that works. Wait for my OK.
4 · Build one step
Do step 1 of the plan only. Then tell me exactly how to test it.
5 · Check it
Explain what you built in plain English and how we know the numbers are right. Work one example by hand.
6 · When it breaks
Here is the exact error: [paste it]. Fix one thing at a time and tell me what you changed.
7 · Wrap up
Write logs/[today's date]_[my name].md with Did / Decided / Open / Next. If we decided something the agent should always know, add it to AGENTS.md. Then commit and push.
Rules that protect you
- No real personal data. No real names, account numbers, or customer records in prompts or code. Use fake data; the website has prompts that make it for you.
- No passwords or API keys in a chat or on GitHub. Keys go in a .env file that is never committed.
- Check before you trust. The agent can be confidently wrong, especially about finance formulas.
- Say how you used AI. Judges score "AI implementation," and that includes being clear about it in your log and demo.
- Stuck for 20 minutes? Ask a mentor, or run My session is going in circles from the website.