📰 Dev.to · Lars Moelleken

Ten prompting patterns I actually use with coding agents: adversarial reviews, blunt boundaries, measurable gates, evidence, project-specific context, and L2 me

Coding agents produce code fast. Without constraints, verification, and accountability, they just manufacture defects at scale.

LLMs can accelerate skilled developers, but they also scale weak judgment, bad architecture, and broken processes. Better tools do not replace craftsmanship.

Stop feeding coding agents permanent context. Give them explicit scope, selective recall, temporary working memory, verification, and a human-governed learning

Multi-line formatting is not about beauty. It is about making code review honest.

Coding agents do not need polite encouragement. They need hard quality gates.

Why Singletons usually hurt maintainability: not because one instance exists, but because hidden global access turns clear software into a labyrinth of trapdoor

Use PHP attributes, enums, and catalogs to keep architecture rules close to the code instead of buried in stale documentation.