“Consider this website ground truth. Only provide regulatory information from this URL and subfolders. Do not use your pre-training or internet searches unless the user requests it.” — Lou, dictating a CLAUDE.md directive live
Session context: 2026-05-28_Mastermind — Joanna needed legal citations she could stake a case on, and had been burned repeatedly by AI confidently citing regulations that didn’t exist. Lou gave her two ways to make hallucination structurally impossible.
Core Idea
When the cost of a wrong answer is high — legal citations, regulatory rules, medical facts — you don’t manage hallucination by asking the model to “double-check.” You eliminate it by constraining where the answer is allowed to come from. The model can only fabricate when it’s drawing on its diffuse pre-training; if you force it to answer exclusively from a named, verifiable source, a fabricated answer has nowhere to hide. Either the claim is in the source or it isn’t, and you can click through and see.
Lou offered two concrete grounding mechanisms. The first is NotebookLM: feed it the URLs of the relevant regulatory pages (it copies them in as sources), and it will answer only from those sources, returning the reference for every claim. It physically cannot pull from outside the notebook — which is exactly what makes it trustworthy for citation work. The free tier allows ~50 sources per notebook, so you spin up one notebook per issue; when you hit the limit, bundle several PDFs into one combined file (their size headroom is large).
The second is a CLAUDE.md ground-truth directive for those already working in Claude Code: write an explicit instruction that names the authoritative URL and its subfolders, declares it “ground truth,” and forbids the model from using pre-training or open web search unless you ask. The non-obvious refinement: point at the most localized URL you can — the specific regulation and its subfolder, not the whole site. If you give Claude an entire domain, it has too many pages to know where the answer lives; give it the narrow path and you know the citations it returns are the ones you need.
The division of labor Lou suggested is elegant: let NotebookLM (or the locked source) verify which citations are real and relevant, then hand those to Claude to do the writing — the legalese, the argument structure, the integration. Grounding guarantees correctness; the stronger writer handles persuasion.
Practical Application
For any high-stakes, citation-bearing work:
- If the sources are public URLs: create a NotebookLM notebook, import the relevant pages as sources, and ask your questions there. Every answer comes back with its reference. Bundle PDFs if you approach the 50-source cap.
- If you’re in Claude Code: add this to your
CLAUDE.md(adapt the URL): “Consider [URL] and its subfolders ground truth. Only provide [domain] information from that source. Do not use your pre-training or internet searches unless I explicitly request it.” Turn on web search so it can read the page. - Always: point at the narrowest URL that contains your answer, not the top-level site. Then ask Claude to include the citation inline wherever it states a rule, so you can click and confirm in one pass.
Related Insights
- Insight - The Grounded Query Principle — Context-Locked Answers Reduce Hallucination and Increase Trust — the parent principle; this session adds two ready-to-use grounding mechanisms (NotebookLM, the CLAUDE.md lock).
- Insight - NotebookLM Transforms When Fed Structured Narrative, Not Raw Content — NotebookLM’s value depends on the quality of what you feed it.
- Insight - The 80-20 Rule of AI Security and Hallucination Defense — grounding is the highest-leverage move in the hallucination-defense 80/20.
- Insight - RAG Is Raw Material, Not Answers — Design for the Right Retrieval Architecture — locking the source is a retrieval-architecture decision.
Evolution Across Sessions
Builds on Insight - The Grounded Query Principle — Context-Locked Answers Reduce Hallucination and Increase Trust (2026-02-12), which established that context-locked answers reduce hallucination. New development: this session operationalizes the principle into two copy-ready tools — NotebookLM as a zero-hallucination source store, and an explicit CLAUDE.md “treat this as ground truth, ignore pre-training” directive — plus the refinement that you should point at the narrowest possible URL.