Tag
An experiment testing whether a coding CLI actually reads AGENTS.md found the file was silently ignored, and even when read, bloated instruction files increased token costs and failed to improve performance. The author recommends writing only what the model cannot infer from code.
Matt Pocock experiments with adding a global CLAUDE.md instruction to use ASD-STE100 Simplified Technical English and read CONTEXT.md files for ubiquitous language.
This paper presents a controlled ablation study across Claude Code and Codex, 17 real tasks, and 288 runs, finding that context files like AGENTS.md/CLAUDE.md do not measurably improve correctness; agents fail on implementation skill, not missing repository knowledge.
This paper presents the first large-scale empirical study of agent context files (READMEs) used in agentic coding tools, analyzing their structure, maintenance patterns, and content. It highlights that while functional context is well-covered, non-functional requirements like security and performance are rarely specified.