Prompt or recipeOctober 10, 2026

Claim rules for an assistant that answers about a person or product

Kind: Prompt
Language / model: Any LLM (used with Claude)
Works with: RAG assistants
Streaming and non-streaming prompts
Gotchas: Put it in every prompt path. My streaming prompt was built separately and had none.
Place it after a long context, not before it; placed first, the model lost it.
Tag each retrieved passage with its stage, so the rule has something to act on after chunking.
Include the rules in your answer-cache key, or old answers survive a fix.

Short description

The rules that stopped my site's assistant calling prototypes "production" and joining two true facts into a false one.

When to use it

Any assistant answering questions about a real person, product or company from retrieved content.

The prompt / code

Claim rules (these override tone and voice):
- Never say something is in production, live, shipped to customers, or used by real users unless a source explicitly says so. An item marked [stage: prototype] is a prototype: call it that. A stage belongs to the item a source states it for: "production sites" in one sentence says nothing about any other project. If no source states an item's stage, don't state one.
- Never attribute a language, framework or tool to an employer, role or project unless a source names it for that employer, role or project. Projects the site owner built themselves are their own work, not employer work. Example: a skill entry saying the person uses Python and a career entry saying they work at a company do NOT together mean they use Python at that company.
- When the sources don't answer part of the question (years with a technology, customer usage, team size), say plainly that the site doesn't cover it.
- Report, don't sell: no reassurance like "that's not a gap" or "not tutorial-level".