Write a System Prompt for Llama 3

Llama 3 uses a structured chat template with special tokens that must be applied correctly for the model to follow instructions reliably. Unlike GPT and Claude where the API handles template formatting transparently, local deployments of Llama 3 (via Ollama, vLLM, or direct inference) require the developer to either use the Hugging Face apply_chat_template method or manually format prompts with the correct special tokens. This example shows the complete formatted prompt for a code review assistant. The system prompt for Llama 3 should be concise and directive, similar to Claude's style. Llama 3 was instruction-tuned with a focus on following direct instructions without elaborate preambles. Unlike GPT models where longer, more elaborate system prompts generally improve results, Llama 3 performs best with tight, action-oriented system prompts under 500 tokens. Avoid vague quality instructions like "be helpful and thorough" — use specific behavioral constraints instead. For Llama 3 specifically, a few formatting tips improve instruction following: use numbered lists for multi-step instructions, avoid nested bullet points deeper than two levels, and put the most important constraint last (Llama tends to prioritize later instructions over earlier ones when they conflict). Test your system prompt with adversarial user inputs before production deployment, as open-source models without safety fine-tuning may be more susceptible to instruction override.

<|begin_of_text|><|start_header_id|>system<|end_header_id|>
You are a senior software engineer conducting code reviews.

Review scope:
1. Identify bugs and logic errors first
2. Note security vulnerabilities (SQL injection, XSS, insecure dependencies)
3. Comment on performance bottlenecks for functions over 20 lines
4. Skip style comments unless they affect readability significantly

Format: Use code blocks for all code references. End each review with a 1-line summary rating: APPROVE / REQUEST CHANGES / NEEDS DISCUSSION.<|eot_id|>
<|start_header_id|>user<|end_header_id|>