Software and platforms

What each platform is, who pays for what, and the course policies that come with working agentically on ANU infrastructure

Participating in this course means having some software installed on your machine and accounts on a handful of ANU and third-party platforms. The onboarding page provides the setup checklist, including versions and the Windows/WSL note. This page explains each platform, who pays for it, and the applicable course policies. Following the documented setup should leave you with nothing to pay.

Windows#

Use WSL2 (recommended) or Git Bash for course development. Use one consistently: tools installed in WSL and Git Bash are separate.

ANU platforms#

• Canvas — the official LMS. Your marks and lecture recordings are published here, and your Claude API key is issued here (assessment submissions go through GitHub, not Canvas)
• Ed Discussion — the course’s one communication channel: announcements, questions and help, and private messages to the teaching team. See the communication page for what goes where

Third-party platforms#

• GitHub — requires a free account. All code submissions go through GitHub repos in the course organisation: private while you build, public from your cutoff onwards (the assessment page covers the flip). You may use a new, anonymous account if you do not want public coursework attached to your usual identity. The course records your handle for marking either way
• Fly.io — no account needed. The deployment target for the full-stack half of the course: every app runs in a course-managed organisation whose billing the course covers, and you get a token scoped to your own app when your first full-stack repo is created, so it can wait until the mid-semester break — your hosting access is the setup, and onboarding step 7 the checklist entry. Don’t sign up or add a payment method for the course

Anthropic is deliberately absent from this list: you don’t need an Anthropic account or subscription to participate. The course manages your Claude access for you by issuing a course-specific API key. Your LLM access covers setup, logging and the usage allocation.

In both halves, your repo determines the live URL of record used for demos and marking. The assessment page specifies the rule.

For all third-party accounts, any paid use is covered by the course provided you follow the documented setup. If something looks like it will cost you money, stop and ask before paying.

Going off piste#

Almost everything on the onboarding list has alternatives: other coding agents, other browsers, other ways of managing runtimes — even other models entirely, whether from a different provider or running locally on your own hardware. None of it is forbidden or discouraged: if you already have a setup you know, use it.

The documented path, though, is the supported one: lectures, demos and tutor help assume it. The course installs its command-line tools with mise, and every template uses it to provision the Node and pnpm versions it was tested against. Using another runtime manager alone does not put you off piste, provided the template’s required versions and commands work; it simply means we will first reproduce a problem with mise. Go off piste when your deviation means the documented template path no longer applies. You then take on the burden of making that setup work — when something breaks in a stack we don’t run, our ability to help you debug it will be limited — and the same assessment requirements apply. You’ll still submit every crit and assignment prototype on time, conforming to the specs and marked in the environment described on the assessment page.

In the final project, choosing the stack is part of the brief: its starter carries no app. The harness, the Fly deploy and the resource limits are supported; tutor help with the stack you choose is best-effort.

Off piste runs all the way to the conscientious objector position: you don’t have to use AI at all. Nothing in the specs requires a single line to be agent-written, and the ethics page makes the convenor’s case for why that stance is a legitimate one to take in a course like this one. What you can’t opt out of is the assessment: the same specs, deadlines and rubric apply, and the process criterion on the assessment page will still ask you to show how the work came together and why you chose to work that way. The same workflow applies: keep the starter’s files and directories, including CLAUDE.md. Working without an agent doesn’t remove the file; it just means you’re the only one reading it. Record there the constraints, standards and feedback loops you actually worked against, and update it as you go; pnpm check:evidence checks the same evidence for everyone.

Permission modes and the YOLO policy#

Running an agent with minimal supervision avoids repeated tool-use confirmations. Claude Code’s permission modes control how much the agent has to stop and ask. The course recommendation is auto mode, which gets you most of the flow benefits while keeping some safety checks in place.

Full “yolo” mode (e.g. claude --dangerously-skip-permissions) removes that last gate. You’re not forbidden from using it, but the course takes no responsibility for anything that happens when you do. If your agent rewrites your repo, deletes uncommitted work, force-pushes over the wrong branch, or burns through your usage allocation in an infinite loop, that’s on you. Commit early, work in a clean branch, and keep an isolated environment for risky runs. Those precautions are your safety net. “The agent deleted my files” is never an acceptable excuse for a missed deadline.