Solve a full course assignment with a coding agent
An instructor found Claude Code could complete an entire assignment including writeup and reflection from the assignment text alone, prompting redesigned assessments and LLM autograding.
Done withClaude Code
- What they did
- In fall 2025, Kästner pointed Claude Code at the assignment text for a Machine Learning in Production homework: extend a small Flask Instagram clone with two AI features. It finished everything without interaction, including the writeup and reflection. He then moved the course to TA check-ins, exams and demo videos, and built LLM-as-a-judge autograding that marks work pass or needs review.
- How it went
- The old assignment became trivial, so it was replaced with a larger task in Zulip, a codebase of over 500k lines. Autograding cut TA grading time by 50 to 80%. He has no evidence on learning outcomes.
- Worth knowing
- Agent-proof designs decay as models improve, so the assignments and countermeasures need rechecking each semester. Calibrating the LLM graders' prompts also takes some tuning.
Try it yourself with Claude Code
Here is the assignment text for my course in [folder or file path]. Act as a student and attempt it end to end, including the writeup and reflection, then tell me which parts you finished with no extra context. Afterward, suggest [3-5] changes to the assignment or grading rubric that would make it harder for an AI to complete without real learning. Save your attempt and suggestions to a new folder and don't edit the original files.
Want help getting this working? Join the Vibe Coding Academy ↗ Run by Alex Finn, who also runs this site.
Source: thelastsoftwareengineer.substack.com · Undated
Related use cases
Turn tickets into merged pull requests
Pulls tasks from GitHub Issues or Linear, then runs Claude Code or Codex in isolated Kubernetes pods to carry each one through to a PR.
Build a thyroid episode warning model from wearable data
Claude Code was given 9.5 years of Apple Watch and Whoop data and built an ML model that backtested as warning of a prior episode before labs confirmed it.
Turn GitHub issues into draft pull requests with long-running agents
The author used Dots for read-heavy long-horizon tasks like reading GitHub issues and making PRs, which were typically poor but gave a starting point.
Build a C compiler with a team of parallel agents
Anthropic's engineering write-up on tasking Opus 4.6 agent teams with building a C compiler.
Resolve a Slack-reported issue with a coding agent
A Codex agent resolved an issue passed to it in about 40 minutes, and an ops agent verified the fix back in Slack.
Turn a website into a mobile app
The author used Replit Agent 4 over about two days and roughly $40 to turn a website into a mobile app with design, preview, and App Store prep, then submitted it for review.
Get 5 like this every morning
The best things people got an AI agent to do, each with the prompt to try it.