E103 - LLM Experience report with Tony Kay
37 snips
Apr 22, 2026 Tony Kay, Clojure engineer and open-source author behind Fulcro and Guardrails, shares his LLM experience. He talks about treating LLMs like junior devs with focused context and stepwise plans. He explains using REPLs, tests, and tooling to prevent hallucinations. He describes Guardrails, signature-based verification, agent orchestration, and how developer roles shift toward design and steering AI.
AI Snips
Chapters
Transcript
Episode notes
LLMs Wrote My Production Unity Plugin
- Tony Kay used ChatGPT to write a complex Unity plugin in Objective-C and Java and, after three or four iterations, he still uses that generated code in production today.
- That first success shaped his trust in LLMs and later led him to have the model fill function bodies and generate tests, multiplying his output three to four times.
Treat The LLM Like A Junior Developer
- Treat the LLM like a junior developer: document the project, write focused specs, and load those files into the model rather than relying on vague prompts.
- Use small, targeted context windows and explicit plan files so the agent stays focused on one incremental step at a time.
Testing And Tool Use Close The Hallucination Loop
- Tool use plus testing closes the hallucination loop: let the model run code (REPL, Node, Python) and write tests so it can verify behavior automatically.
- Verification via executable tests transforms token prediction into trustworthy development autonomy.
