Theo Browne explains how he replaced generic global coding instructions with a smaller set of preferences grounded in his actual work. He uses conversational guidance to establish tone, scope discipline, code-quality expectations and safety boundaries, while keeping repository-specific knowledge in local AGENTS.md files.
Theo Browne builds focused skills for recurring workflows such as filing and monitoring pull requests, uploading review artifacts and communicating through readable HTML. He emphasizes that a skill description should primarily contain the trigger language that helps an agent decide when to load the detailed instructions.
Theo Browne also audits prior agent sessions to find repeated failures by model and harness, then converts those observations into concise safeguards. Examples include preventing destructive process management, limiting scope creep, covering every product surface, preserving reverse actions and separating maintainer documentation from user documentation.
The broader lesson is to treat instruction files as an evolving collaboration layer rather than a template to copy. Theo Browne recommends studying where agents misunderstand intent, waste tool calls or communicate poorly, then writing the smallest instruction or skill that addresses that specific pattern.
Watch on YouTube
