You already know a hallucination is the model stating something false as fluently as something true, and that you can't spot one by tone. That was the warning. This node is the practical follow-up: you can't switch hallucinations off, but you can push their rate down and make the ones that slip through easier to catch. Three habits do most of the work.
A model answering from training data is recalling patterns, and recall is where it invents. The moment the real information is sitting in the prompt, the job changes from "remember this" to "read this and answer", which it is far better at.
So when an answer depends on a specific document, a config file, an error log, or a page of docs, paste the relevant part into the prompt and ask the model to answer from that text. This is called grounding, and it is the single most effective move you have. A later topic is entirely about doing it well.
Left to itself a model treats every question as answerable, so a gap in its knowledge gets filled with a confident guess. You can change that with one line in the prompt: tell it to say it doesn't know when it isn't sure, and to answer only from the information you gave it.
claude -p "Using only the text below, answer the question. If the text does not contain the answer, say you don't know. ..."
It won't be perfect, but an explicit out makes a guess less likely than an unspoken expectation that it must always produce an answer.
For claims that carry a cost if they're wrong (a command you'll run as root, a legal or medical detail, a figure going into a report) ask the model to cite where each claim comes from, then open the source yourself. A citation is useful only when you verify it. The model knows what a plausible URL or quote looks like, so it can produce one for a fact it invented. A link that 404s or a quote you can't find in the page is the tell.
Grounding, an explicit "I don't know", and verifying costly claims all point the same way: keep your hand on the wheel. None of them make the output trustworthy by default. They make it faster to produce and easier to check, which is the whole point of using AI for real work.
Your lab setup
This VM comes preconfigured with your AI keys - just type claude to get started. Your key and URL are already set as $OPENAI_BASE_URL and $OPENAI_API_KEY.