prompt
A prompt is what a person says or types to get an AI to do something. It is the question, instruction, or idea you hand over. People think of it as starting a conversation, like asking a friend for help.
What a model may hear
- text injection security and red-teaming
- treat the input as a possible attack vector and filter or block it
- model input token sequence machine learning inference
- tokenize, truncate, or pad the text to fit context-window limits
- shell command prefix operating systems and CLI tools
- expect executable syntax, flags, or pipe chains and may try to run code
- training example fine-tuning and datasets
- match the input against known prompt templates or few-shot patterns
- diffusion image prompt image generation
- parse for weight syntax, style tokens, or negative prompts
Where people and models part ways
“Write me a prompt for a story.”
Meant: Give me a good idea to tell an AI.
May be taken as: Return a formatted JSON prompt object or system prompt template.
Say instead: “Give me a fun idea I can type into a chatbot to get a short story.”
“Prompt me when you're ready.”
Meant: Let me know when I should continue.
May be taken as: Generate a shell script prompt or stall waiting for structured input.
Say instead: “Tell me when you're ready for me to go on.”
“I need a better prompt.”
Meant: My question was unclear and got a bad answer.
May be taken as: Suggest prompt-engineering frameworks or injection-resistant templates.
Say instead: “Help me rephrase my question so the AI understands me better.”
“No prompt needed.”
Meant: I do not want to give instructions right now.
May be taken as: Assume zero-shot inference or bypass safety layers.
Say instead: “I will just chat without a specific task in mind.”
Tips
- Say 'my question' or 'what I typed' if you mean ordinary words, not a engineered prompt.
- Avoid shell-like syntax such as pipes or flags unless you want code.
- If you want a simple answer, say 'answer simply' rather than 'zero-shot prompt.'
- Do not use prompt-engineering terms like 'system prompt' or 'temperature' unless you know the model exposes them.
- Treat the word as ordinary speech when talking to people, and as structured input only when talking to the model about its own behavior.
Often confused with
- command
- Implies execution, not conversation.
- query
- Suggests a database lookup, not open-ended generation.
- instruction
- More one-way and prescriptive.
- input
- Broader, includes non-text data.
- priming
- Psychological framing, not technical token setup.