LLMs produce statistically likely answers, not correct ones — a critical distinction for software engineering. Drawing on Jonathan Snyder's paper 'Triangulating for Truthiness,' this piece outlines practical principles for maintaining engineering judgment when using AI coding tools. Key ideas include: codifying architectural specs as explicit artifacts the LLM must honor, treating every suggestion as a hypothesis requiring validation, recognizing 'cycle spinning' as a trap, applying the five whys methodology in stages rather than jumping to conclusions, using automated tests as an 'AI Andon cord' to halt work when AI-generated changes exceed safe evaluation scope, and actively managing context drift across long sessions. The overarching philosophy borrows from Intent-Based Leadership — express objectives and constraints rather than issuing task-level commands, keeping the human as the accountable decision-maker.
