Skip to main content

🧾 Components of a Prompt in Prompt Engineering

A prompt's form depends on the task you are giving to a model. When exploring prompt engineering, you will often see prompts that include some or all of the following elements:

🔹 Instructions​

  • This is the task for the foundation model (FM) to perform.
  • Provides a description or directive on how the model should behave.

🔹 Context​

  • Offers external information to guide the model.
  • Helps improve relevance and accuracy.

🔹 Input Data​

  • The user-provided content that needs processing or response.
  • For example: text to summarize, a question to answer, etc.

🔹 Output Indicator​

  • Specifies the desired format or type of output.
  • Helps the model understand how to structure the response.

🧭 Model Latent Space

  • Represents the “hidden” knowledge stored in a trained model
  • It's a statistical database of patterns and relationships learned during training
  • When you prompt an LLM, you're navigating this latent space to retrieve relevant knowledge

Example:​

  • A scuba vacation recommendation model might encode:
  • Destinations
  • Dive depths
  • Visibility
  • Weather conditions
  • Even if a user asks, "Where can I snorkel with manatees?", the model can infer an answer using its latent understanding of related concepts.

❌ Negative Prompting?

Negative prompting is a technique in prompt engineering where you explicitly instruct the model to avoid certain behaviors, topics, formats, or types of content in its response.

🧾 Why Use Negative Prompting?​

LLMs are flexible, but sometimes they:

  • Drift off-topic
  • Make assumptions
  • Include unwanted content (e.g., code, hallucinations, opinions)

Negative prompting helps steer the model away from these undesired outputs.

✅ Examples​

Content Restriction​

Write a summary of this article, but do NOT include any opinions or assumptions.

Format Control​

Explain this concept in plain language. Do not use technical jargon.

Avoid Hallucination​

Answer only based on the context provided. Do not make up facts.

Prevent Repetition or Verbosity​

Give a short response, and do not repeat any part of the input.

🔧 Use Cases for Negative Prompting​

Use CaseNegative Prompt Example
📱 Chatbots"Do not ask the user for personal data."
⚖️ Legal Summarization"Avoid giving legal advice."
🎓 Education"Do not include the final answer, only explain the steps."
🏥 Healthcare"Do not make a diagnosis or suggest medication."