Chapter 8 Avoiding Hallucinations . A grounded QA template: extract the most relevant quote into <scratchpad , judge whether it answers the question, and only then write a brief nu
From the Anthropic cookbook's building moderation filter.ipynb . A production style moderation prompt that uses explicit ALLOW/BLOCK guidelines, a <thinking scratchpad, and an <out
RAG system prompt from the embeddings QA recipe. The closing sentence is what stops hallucinations on questions outside the retrieved articles. Paired role message: You answer ques
One line system prompt from the function calling recipe. It is responsible for the model's habit of asking clarifying questions rather than fabricating function arguments. Pair it
Chapter 4 Separating Data and Instructions . Compares the un tagged form 'Yo Claude. {EMAIL} < Make this email more polite' with the safer tagged form below. The XML wrap is what s
Opening sections of Perplexity's leaked system prompt. Strict on inline numeric citations, markdown structure, no moralization, and no "based on search results" hedging. Source: ht
Opening sections of Anthropic's officially published system prompt for Claude Sonnet 4.5 release notes, Sept 2025 . Notable for prose first formatting guidance, sparse bullet usage
Opening section of Devin's leaked system prompt. Establishes Devin's identity as an autonomous SWE, defines when to communicate with the user, and lays out the planning vs standard
The most famous prompt from the awesome chatgpt prompts repo 1,800+ contributors . Turns the model into a deterministic Linux terminal emulator that only outputs what the shell wou
From the official Anthropic prompt engineering guide for the Claude 4.7 family. Two single line additions that calibrate verbosity in either direction useful as a system prompt fra
Opening of the Bolt.new system prompt. Documents WebContainer's browser bound Node.js runtime, the lack of native binaries, pip, and git, and the bolt file modifications diff forma
The few shot variant of Anthropic's moderation cookbook. Demonstrates that, for short labels, well chosen positive and negative examples often beat chain of thought. Source: https:
From the cookbook's using citations.ipynb . Pairs the Claude Citations API with a short, focused system prompt to ground answers in help center documents. Cited spans are returned
Microsoft's documented chatbot template. The two line instruction block is what keeps the assistant from answering off topic questions, and the IRS fallback URL prevents hallucinat
The most pulled RAG prompt on LangChain Hub. Three sentence answer cap plus an explicit "say you don't know" clause makes it a safe default for production RAG. Source: https://smit