Prompting Multimodal Models for Image‑Text Tasks

Effective multimodal prompting combines a clear system instruction with a few example pairs. 1. Define the system role → "You are an assistant that describes images concisely." 2. Provide a few‑shot example: ``` User: Assistant: A tabby cat lounging on a windowsill. ``` 3. Use explicit separators → `---` between image token and text. 4. Keep the user query short and focused. ``` User: Explain the flow in two sentences. ``` 5. If the model supports temperature control, set `temperature=0.2` for deterministic captions. ```bash curl -X POST ... -d '{"temp":0.2}' ``` **Gotcha:** Avoid mixing multiple images in one prompt unless the model explicitly supports batch processing; it can confuse the output.