In this technical report, I document lessons learned from building an LLM-centric prototype for a science-intensive astrophysics application.
In this experiment, Claude acted as the “chief scientist” and system architect; Codex CLI served as the implementation layer, ingesting Claude’s prompts and generating code; and I performed validation by comparing model outputs against published astrophysical observations.
Result: the LLMs were effective at generating ideas and code, but their ability to produce scientifically valid astrophysical results was limited.
Yes, it’s obvious that LLMs cannot do astrophysics at a professional level. The point here was to demonstrate it empirically—using Claude for design and testing logic, and Codex for implementation. The tooling was effective; the science was not.
See Also on Research Gate:
