Seven AGENTS.md Rules for Vibe-Coding Agents

1. Never edit, skip, or delete a test to make it pass. For every test you write, confirm it would fail if the code under test did nothing. If it would still pass, rewrite it. A green test suite is the easiest thing for an agent to fake. 2. Before adding any dependency, confirm it exists on the official registry, who maintains it, and when it last shipped. Show me the link. Never guess a package name. Agents repeat the same fake package names, and attackers can register them. 3. Before using any library's API, check the installed version and read its actual docs or source. Never write code from memory of an older version. Agents write code for the version they were trained on. Yours could be newer or older. 4. Never say a task is done until you've proven it against the real thing. Run it and show me the actual output, not a summary. 'It compiles' is not proof. If you can't run it, say so. When agents grade their own work, they rarely fail themselves. 5. If two fixes fail the same way, stop. The shared assumption is probably wrong. Question it before writing another fix. Never retry a theory you already ruled out. Agents retry the same wrong theory with more confidence every time. 6. Only change what a task requires. If you notice other issues, first list them instead of immediately fixing them. The bugs you didn't ask for come from the fixes you didn't ask for. 7. Never delete untracked files or data, rewrite git history, or force-push without showing me the exact command and waiting for approval. One command can erase a day of work, and agents run them casually.