An engineering manager argues that AI coding tools need better context management—not just bigger models—to reduce hallucination and maintain developer control.

**What the author observed:** AI coding models struggle not because they lack capability, but because they lack sufficient context about the developer's situation—the codebase structure, team conventions, earlier conversations, and the specific problem being solved. Without that context, models guess and produce well-formatted answers to problems that were never asked.

**How the author structures the solution:** Using a "harness" (the controls and checks around an AI model, separate from the model itself), implemented as plain-text specification files in a `.plans/` directory. The harness includes guides (like specs that steer the model before it acts) and sensors (like tests that flag problems after the fact), split between computational checks (deterministic—type checkers, linters) and inferential checks (AI evaluating AI's output). Four documented skills—product-requirements, feature-development, code-review, and testing-pyramid—formalize the flow from idea to implementation.

**Why it matters:** This approach keeps quality "left"—catching problems in the spec rather than the pull request, or in the PR rather than production—which is cheaper and gives developers explicit control over what the agent sees and does, rather than relying on the model to guess correctly.