Agent Four Principles

Problem: LLM code generation tends to be fast but noisy

LLMs often make hidden assumptions, add unnecessary abstractions, refactor adjacent code without being asked, and treat "it runs" as completion. This skill turns those failure modes into a set of conservative engineering rules.

How it works: four principles constrain the coding loop

• Think before coding: state assumptions explicitly, surface ambiguous interpretations, and push back when needed.
• Simplicity first: deliver only the minimal code required; avoid one-off abstractions, unrequested configurability, and error handling for impossible cases.
• Surgical changes: touch only lines that map to the request, match existing style, and flag unrelated dead code instead of deleting it.
• Goal-driven execution: rewrite weak tasks into verifiable steps, such as "write a failing repro test, then make it pass."

It is useful when writing new code, reviewing or refactoring, handling complex decisions, or responding to a request for code-quality review. It does not replace tests or human review; it replaces vague acceptance criteria with checkable checkpoints. For trivial tasks, the conservative bias may still require judgment.