Hard Prompt
A hard prompt is ordinary natural-language text (and special tokens) fed to a model—the human-readable prompt string. Unlike a soft prompt, every token maps to the vocabulary; you can copy, version, and audit it as text.
Why it matters
- Default interface for chat, RAG, and agents
- Portable across models if you adapt the chat template
- Debuggable: you can read what the model actually saw
How it shows up
Prompt engineering edits hard prompts: instructions, few-shot exemplars, tool schemas. Prompt tuning instead learns soft vectors while keeping the base weights frozen. Hard prompts win on interpretability and zero-shot reuse; soft prompts can beat them on a fixed task when you have labeled data to optimize against.
Use hard prompts when you need reviewable control; reach for soft prompts or adapters when you are optimizing a stable task end-to-end.