Dan Luu warns passive LLM coding fails

Acting as an unthinking relay between LLM errors and prompt retries is a dead-end strategy that conflates raw iteration speed with genuine engineering competence.

• –The "meat proxy" trap: If a task can be solved simply by feeding error messages back into an LLM until tests pass, an automated script can do it cheaper and faster than a salaried engineer.
• –Invisible technical debt: Unsupervised AI code frequently overfits to metrics, generates redundant abstractions, and fails catastrophically on out-of-distribution problems like complex build systems.
• –The pre-prompting imperative: Engineering leverage comes from architectural planning, domain modeling, and strict verification before invoking agents, rather than letting models wander into compounding regressions.
• –The evaluation gap: Hype-driven assertions that coding is solved fall apart under real-world inspection, where unsupervised agent outputs routinely inflate diff sizes and mask fragile failure modes.