The team friction Teams start relying on LLMs for everything, including validating code correctness, checking style, or enforcing architecture rules. But LLMs are probabilistic - they occasionally hallucinate, drift from instructions, or miss edge cases. This leads to inconsistent enforcement of team standards and erodes trust in the automated pipeline.
The pattern The team aggressively codifies rules. When an LLM correctly catches a bug or enforces a pattern in review, the team extracts that rule into a deterministic tool—a linter, a unit test, a static analysis check, or a custom script. They explicitly offload the “known knowns” to rigid, non-AI tooling.
Why it works with AI It frees up the AI’s context window and reasoning capabilities for high-level problem solving, rather than wasting tokens and time on things a linter can do perfectly, cheaply, and instantly. It creates a robust safety net that lets the AI agent move fast without breaking things.
Watch-outs Don’t try to make everything deterministic on day one. Let the AI find the patterns and edge cases first, then codify them once they prove valuable. Avoid writing custom linting rules for things that don’t actually matter to the team’s velocity.
Related Consistent, excellent codebase · Cleanup time