How to use Claude to prototype subway predictions
Claude Code drafted image collection, labeling and PyTorch prediction tools for a subway-status app, but the author needed substantial manual fixes.
news.ycombinator.com
Prompt to try
Editorial adaptation, not the source author’s exact prompt. Replace bracketed inputs and review tool permissions first.
Build a prototype for [transit system] using publicly available [circuit-diagram images]. Draft a background downloader, a tkinter labeling tool and a PyTorch prediction script. Make outliers and GPU configuration visible for my inspection.
Reach people exploring AI agents
See placements →What happened
The author built and shared an MIT-licensed subway-status project, describing substantial manual corrections and a missed GPU configuration; no accuracy figures are given.
Reported by MrEricSir on Hacker NewsHow the agent did it
- Had Claude Code write a background subway-diagram downloader.
- Had it build a tkinter image-labeling tool and manually adjusted it.
- Used labeled images with a Claude-written PyTorch prediction script.
Where a person stepped in
The author manually tweaked the labeler and prediction script, reconsidered outliers and discovered GPU support was not enabled.
A person guided the agent closely as it worked.
A San Francisco subway outage prompted the author to build a web app intended to help answer whether to take the subway or a bus. They chose to use Claude Code extensively as a way to learn new tools. The work began with a background script that downloaded publicly available images of the subway's real-time circuit diagram from sfmunicentral.com.
Next, the author had Claude Code build an image-labeling tool using tkinter. That tool required substantial manual adjustment before labeling could begin. The author says writing it from scratch would have saved time, despite tkinter seeming appropriate for the task.
The next stage used the labeled images for predictions in PyTorch. Claude produced the initial script quickly, but the author again needed manual changes and reconsidered images marked as outliers. They later discovered that Claude had not enabled PyTorch's GPU support. The post shares the project source under an MIT license, but provides no prediction-accuracy results. Who it's for: developers experimenting with image-based transit prototypes who can inspect generated data tools and machine-learning configuration themselves.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Hacker News · MrEricSir
Frequently asked questions
Where did the input images come from?
The author used publicly available real-time subway circuit-diagram images from sfmunicentral.com.
What needed manual correction?
The tkinter labeler required extensive tweaking. The prediction script also needed adjustments, and PyTorch GPU support had not been enabled.
More details
Remaining steps
- Manually tweaked predictions and discovered missing GPU support.
About this account
Claude Code generated initial scripts for collection, labeling and prediction; the author performed substantial manual adjustment.
Platform: Hacker News. Attribution: https://news.ycombinator.com/item?id=46637283. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: MrEricSir on Hacker News
Source published 15 Jan 2026
Source checked 14 Oct 2026
Tools and topics
Claude Code · tkinter · PyTorch · sfmunicentral.com
Agent workflows
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.