coding-guidelines

To prevent common LLM coding mistakes, apply behavioral guardrails that enforce surfacing assumptions, favor minimal changes over speculative features, and require verifiable success criteria before committing changes.

AI generated code often includes unnecessary abstractions because LLMs lack inherent constraints against speculative features. Enforcing a simplicity-first guardrail ensures the model prefers minimal code that directly solves the task and avoids unrelated refactors.

The best way to ensure LLM code edits are verifiable is to define strict success criteria beforehand and enforce surgical changes. This means making only request-traceable edits and removing only artifacts created by the specific change.

To make an LLM ask clarifying questions before implementing code changes, apply a think-before-coding guardrail that forces the model to surface assumptions, present tradeoffs, and propose a concise plan prior to any code modification.

Yes, you can use coding guidelines for bug fixes where conservative edits are required. They enforce goal-driven execution by defining verifiable success criteria and iterating until those checks pass without altering unrelated code.

You should not use automated refactoring with an LLM when a task lacks clear verifiable success criteria, as the model may introduce inappropriate refactors, incorrect assumptions, or overcomplex solutions instead of surgical changes.