Prompting Llama 4
How to prompt Meta's Llama 4 (Scout, Maverick), the most widely-deployed open-weight family in enterprise. Read the shared Prompting guides principles first; this guide carries only what is distinctive to Llama 4. This is reference for when you target, self-host, or evaluate Llama 4.
When to reach for it
Llama 4 is the default "safe" open-weight choice — broadly supported across inference stacks, with Scout's large context window suiting long-document work. Reach for it when enterprise deployment breadth, tooling maturity, or an established open baseline matter more than topping the open-model leaderboard.
Prompting principles
- Match the exact chat template. Llama 4 uses its own header tokens (
<|begin_of_text|>,<|start_header_id|>,<|eot_id|>, …) — distinct from Llama 2/3. Let the framework (Ollama, vLLM, llama.cpp) apply the template, and if hand-building, keep the required newlines. Multimodal input needs<|image_start|>/<|image_end|>boundaries per image. - Keep the system prompt short. 2–4 sentences: define the role, output format, and behavioural constraints, then stop. Long system prompts make it lose focus on later instructions — push extra context into the user message.
- Use a system prompt deliberately. A good one measurably reduces false refusals and "preachy"/templated phrasing, and improves conversationality and formatting — Meta calls Llama 4 highly steerable.
- Few-shot + prefilled structure. For structured output, give one or two input/output examples and prefill the opening token (
{for JSON,<summary>for XML) to lock the format. Tag sections (<task>,<rules>) to stop instruction overlap. - Frame long documents first. At long context, precede a large document block with a brief statement of what it is, its format, and what you need — before the content itself.
- Layer safety at the system level. Don't rely on prompts alone for safety; pair with Llama Guard 4 as a system-level guardrail.