We read every Llama 3.x prompt guide so you don't have to.

What Llama 3.x actually rewards

We pulled this from Meta's official guidance and what works in production. The short version:

• →Use exact chat template (or HuggingFace chat_template abstraction).
• →Set a system prompt.
• →Few-shot helps — add examples.
• →Specify format ('Respond with the integer number only').

Common mistakes we fix automatically

• Avoid
  Don't omit special tokens.
• Avoid
  Don't expect tool-call + multi-turn chat to be reliable on 8B (use larger sizes).

Frequently asked questions

Which versions of Llama 3.x does this support?

We support llama-3, llama-3.1, llama-3.2, llama-3.3. We apply the prompt patterns Meta recommends for each, so the rewrite is tuned to the version you're using.

Is my prompt stored or used for training?

No. Prompts are sent to the rewriter, scored, returned, and discarded. We don't train on them and we don't keep them around.

Do I need to know prompt engineering to use this?

Nope. That's the point. Paste what you have, click Rewrite, get back a version that follows Meta's official guidance.

What makes this different from Llama 3.x's own "improve prompt" feature?

Built-in optimizers use the model's own preferences. Ours is built on Meta's official documentation and patterns that consistently produce better results in production. Llama 3.x works best with prompts in the 100-3000 token range, and we keep rewrites inside that window.