# Instructions name only what exists **The rule.** An instruction file (CLAUDE.md, AGENTS.md, a skill, a memory note) names only paths, links, commands and options that exist now. Anything it names that can be checked is checked mechanically, and a stale one fails the build rather than waiting for an agent to trip on it. Limits and numbers live in the code; the instruction points at the code instead of restating them. **Why.** Agents follow the instruction file over the code. A published study of 40 projects found a stale CLAUDE.md or AGENTS.md instruction followed about 90% of the time, against about 4% for what the code actually did ([SmartScope](https://smartscope.blog/en/blog/claude-md-agents-md-stale-instructions-override-code-2026)). Ours drifted the same way: a guard kept protecting a path deleted four months earlier, and a posting skill kept a daily limit after the code that enforced it changed. **Enforced by** - `rstack/tools/instruction-refs`: resolves every markdown link, every repo path in backticks, every `rh` subcommand and action and every `rh` option named in the given files, and with `--local` every `~/` path on this machine. A skill's own files resolve against its folder first. - [test](../../tests/test_instruction_refs.py): the repo's own AGENTS.md, CLAUDE.md, README, docs, skills and rstack pages pass it on every run of the suite. **How to apply.** Run it with `--local` over the instruction files and memory folders that live outside a repo, as part of whatever pass already reorganizes them. A finding is either a stale line to fix, or a sentence that names a path on purpose because it is gone; reword that one so it says so. Addy Osmani's agent-skills enforces the same thing in CI (`validate-reference-links`, `validate-commands`, `validate-artifact-paths`).