How to use Cursor to reproduce classifier research
Using Cursor with GPT5-Codex, the author built a classifier grounded in local research materials, compared results, fixed a training-data mismatch and ported the code to Go.
news.ycombinator.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Reproduce [classifier paper] using [local paper text] and [reference repository]. Cite the supplied paper when explaining design decisions, create result-comparison scripts, check training-data construction and tune parameters from observed results. Document discrepancies before extending the simulator.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source poster / publisher
- frikk
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- The author reported a classifier, comparison scripts, a corrected training-data mismatch, a parallel Go port with a claimed tenfold speedup and a simulator working on its first attempt.
- Human involvement
- The author supplied the paper and source, discussed design choices and tuned parameters using empirical outcomes.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Cloned the reference repository and copied the research paper into a local text file.
- Generated comparison scripts and used paper findings and empirical results to guide fixes and tuning.
- Identified the mismatch between random training data and the paper's solution-space permutation.
The author built an experimental cellular automata classifier in one evening using Cursor with GPT5-Codex. A central step was placing both the original repository's source code and the research paper inside the local project. They cloned the repository and copied the paper into a text file, giving the model references for design questions.
With those materials available, the model created scripts to compare results against the paper, fixed bugs and helped tune parameters based on empirical outcomes. The author highlighted one important discovery: their training data was randomly generated, whereas the paper used a permutation over the entire solution space. The model identified this mismatch as a critical bug.
They also ported the code to Go and parallelized it, reporting a tenfold processing speedup without supplying measurement details. Finally, the author requested a simulator using invented tests involving hypothetical sensors and unimplemented data. They reported that it worked on the first attempt without changing the original engine implementation.
Who it's for: Developers reproducing research with accessible papers and reference code.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Hacker News · frikk
Frequently asked questions
What local evidence helped ground the model's work?
The author cloned the original repository and copied the research paper into a local text file, enabling references to the paper during design discussions.
What training-data problem did the model discover?
The implementation used randomly generated training data, while the paper used a permutation over the whole solution space. The author described this mismatch as a critical bug.
More details
Remaining steps
- Ported the implementation to Go, parallelized it and requested an additional simulator.
About this account
The model generated tests, fixed bugs and built a simulator while the author supplied references, questions and experimental requirements.
Platform: Hacker News. Attribution: https://news.ycombinator.com/item?id=45926874. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: frikk on Hacker News
Source published
Added to Agents In Use
Source checked
Tools and topics
Cursor · GPT5-Codex · Go
Research reproduction · Cellular automata · Go
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.