reality-check

# Reality Check Validate AI-generated technical guidance before the user wastes time acting on it. Modern models produce recommendations that are fluent, confident, and structurally perfect — and quietly riddled with invented package names, nonexistent repos, made-up config keys, plausible-but-wrong CLI flags, and "industry-standard patterns" that no one has ever used. On top of that, even the *accurate* parts are often overkill: complexity bolted on for its own sake, reinventing capabilities the user already has. The user's ask is almost always some version of *"is this real, and is it worth doing?"* Answer both. Be brutally honest — a soft, agreeable review is worse than useless here, because the whole point is to catch what the original generator got wrong. ## Core principle: verify, don't vibe The failure mode of this skill is reviewing claims from memory and *sounding* confident. Don't. Memory is exactly what hallucinated a wrong package name in the first place, and your training data goes stale. **For every checkable claim, actually search.** A claim you "recognize" is not a verified claim — version numbers, repo ownership, CLI flags, and tool capabilities all drift. When in doubt, look it up; when a search is cheap and the claim is load-bearing, look it up anyway. ## Workflow ### 1. Frame the artifact and extract discrete claims Identify what you're validating (a plan, a tool list, a config, a step-by-step setup). Then break it into **atomic, individually-checka...