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 also use "prompt" for the moment the cursor blinks and waits for them to type.
What a model may hear
- injection attack vector AI safety and security literature
- treat user input as potentially hostile; may refuse, sanitize, or wrap the text in defensive system instructions
- token sequence prefix transformer model internals and inference code
- interpret the word as the fixed text that conditions all following generation, including hidden system prompts and template fragments
- engineering discipline productivity guides and model documentation
- assume the user wants structured techniques such as chain-of-thought, few-shot examples, or role assignment
- shell command prompt operating systems and CLIs
- expect a command-line interface context where "prompt" means the $ or > symbol and the state of the terminal session
Where people and models part ways
“Fix my prompt.”
Meant: Make my question clearer so I get a better answer.
May be taken as: Rewrite the system prompt or hidden instructions, possibly overriding safety filters or changing model behavior globally.
Say instead: “Rewrite my question so the model understands me better.”
“I got prompt injected.”
Meant: Someone tricked the AI with clever wording.
May be taken as: Assume a formal security incident requiring logging, escalation, or model retraining rather than a single misleading query.
Say instead: “Someone used tricky wording to make the AI ignore its instructions.”
“Prompt it again.”
Meant: Ask the same thing a second time.
May be taken as: Append the same token sequence to the context window, causing repetition loops or context overflow.
Say instead: “Ask the model the same question again in a fresh conversation.”
“My prompt is too long.”
Meant: I wrote too many words for the AI to handle well.
May be taken as: Measure against the model's token limit and truncate or summarize automatically, losing the user's intended content.
Say instead: “My message is too long; please help me shorten it while keeping the key points.”
Tips
- Say 'my question' or 'what I typed' instead of 'my prompt' when you just mean ordinary words you said to the model.
- If you want structured output, say so directly rather than asking the model to 'engineer the prompt.'
- Start a new chat for a new topic; 're-prompting' in the same thread adds to context the model must sort through.
- When something feels like a trick, describe the trick in plain language rather than using 'prompt injection' as shorthand.
- If a model seems to follow hidden rules you did not write, ask what instructions it is following rather than assuming your 'prompt' controls everything.
Often confused with
- command
- Implies obedience and exact execution, not conversation.
- query
- Suggests a database lookup, not open-ended generation.
- input
- Broader and more technical; loses the idea of asking.
- instruction
- Stronger authority; models may treat it as a system-level override.
- priming
- Psychology term; in models it means biasing via earlier text, not the current request.