The editor war between Emacs and vi users has run since at least 1985, and it was one of the original holy wars on Usenet. Developers now argue about coding agents the way they argued about editors.

An editor changes how a person types code, and an agent writes the code. That moves the skill from keystrokes to instructions.

The CLAUDE.md guide in this topic tests four repositories and finds that Claude Code prefers CLAUDE.md by default and falls back to AGENTS.md. What the agent knows about a project is what a person wrote in those files.

A coding agent is configured through files a person writes, so getting good output from one is a documentation job.

Instruction files

Delegation

Memory

Choosing a model or tool

What the code is worth

What does a person have to write for a coding agent?

A person writes the instruction file the agent loads and the prompt file for each subagent. The CLAUDE.md guide tests which file each agent loads, and Claude Code subagents start without the conversation history, so the prompt file has to carry the rules.

A coding agent is one agent loop with a harness around it, which the AI agent architecture topic takes apart.

Claude Code's memory is one case of a wider design, and the AI agent memory topic covers the write and read paths behind it.

Instruction files are documentation, and the technical documentation topic covers the page types they sit beside.