- Identify what LLM (brain e.g. gpt4o-mini) and what TTS (voice, e.g ElevenLabs or Cartesia) you want to use. (Ask if unsure for your specific case study info@anam.ai)
- Open up ChatGPT and copy the code. Make sure you replace the areas in red with what model you are using or specific citation examples.
<aside>
💬
I want you to create a large system prompt for an LLM, to be used in [INSERT MODEL NAME, e.g., GPT-4.1 mini, GPT-5-turbo] with [INSERT VOICE MODEL, e.g., ElevenLabs Conversational Voice Mode] consisting of.
RULES
- Do not invent content. Only use what I provide in the uploaded documents.
- The system prompt must be structured with labeled sections:
- [GOAL] — identity, purpose, scope, mandate
- [COMMUNICATION STYLE] — voice delivery, tone, uncertainty handling, citation rules
- [RESPONSE GUIDELINES] — procedural compliance, step-by-step logic, error handling
- [KNOWLEDGE BASE] — segmented by document, expanded with exact section numbers and titles
- In the [KNOWLEDGE BASE], expand the largest document in the set the most, to reflect its weight (e.g., if one is 22 pages, it should dominate this section).
- Always cite document + section: e.g.
“citation format”
- Do not summarize too loosely. Expand into usable procedural detail. Avoid copy-pasting entire long passages verbatim.
- Voice delivery rules must follow
VOICE SYSTEM best practices (e.g., ElevenLabs: <https://elevenlabs.io/docs/conversational-ai/best-practices/prompting-guide>).
- Balance content to reach the requested size:
- ~4,000 tokens → deliver in 2 parts
- ~6,000 tokens → deliver in 3 parts
- If user does not want segmentation, deliver in 1 Markdown code block.
OUTPUT FORMAT
- Always ask first how many tokens you want the prompt to be.
- Always provide output in Markdown code blocks (```) so I can copy it directly.
- Segment into parts depending on the requested size:
- 2 parts (~4k tokens):
- Part 1: GOAL + COMMUNICATION STYLE + RESPONSE GUIDELINES
- Part 2: KNOWLEDGE BASE
- 3 parts (~6k tokens):
- Part 1: GOAL + COMMUNICATION STYLE
- Part 2: RESPONSE GUIDELINES
- Part 3: KNOWLEDGE BASE (expanded heavily)
- Clearly label parts.
CONTEXT
The uploaded documents are the only knowledge base to use.
If content is missing, exclude it — do not hallucinate or fill gaps.
</aside>
- Upload this alongside your document in chatGPT.