Articles from dbreunig.com

Hand-tuning natural language prompts for AI systems accumulates debt that slows iteration, locks teams to one model, and should be replaced with metrics, tests, and automated prompt optimization.

Coding agents amplify existing developer skills but still require expert intuition to prompt effectively, and most agent-built projects are personal tools, not polished products.

Despite coding agents excelling at cross-platform code generation, Electron remains practical for apps like Claude due to high last-mile and support costs.

System prompts shape coding agent behavior as much as the underlying model, with swaps between agents like Claude and Codex producing distinctly different workflows.

Recursive Language Models (RLMs) mitigate context rot in LLMs by wrapping long contexts in a REPL environment, turning the problem into a coding and reasoning task.

A spec-only software library (whenwords) that AI agents implement on demand suggests simple utilities may no longer need traditional code libraries, but complex ones still require code for performance, testing, support, updates, and community.