E103 - LLM Experience report with Tony Kay

Software development is shifting from manual algorithm implementation to system design and AI orchestration. Tony Kay, a Clojure developer and open-source author, advocates for a spec-driven approach where LLMs act as junior developers guided by precise, modular instructions. Maintaining context window efficiency—ideally below 50%—is critical for performance, achieved by offloading tasks to sub-agents and using "skill files" that encapsulate domain-specific knowledge. Verification remains the primary defense against hallucinations; by integrating LLMs with REPLs and automated test suites, developers can ensure code reliability without manual oversight. This evolution renders low-level coding skills secondary to the ability to steer AI agents, define clear constraints, and build deterministic, testable systems. Ultimately, the future of engineering lies in creating robust, self-verifying architectures that allow developers to focus on high-level logic rather than syntax or boilerplate.