Note 2—Prompt Engineering for Developer

Topic: Guideline for Prompting

Photo by Matt Palmer on Unsplash

Prologue

This note emphasizes on 2 key principles and 6 tactics for how to write prompts. With proper prompts, it will increase the chance to get the results that you want.

Principle 1: Write Clear and Specific Instructions

• Tactic 1: Use delimiters
• Tactic 2: Ask for structured output
• Tactic 3: Ask the Model to Check whether conditions are satisfied
• Tactic 4: “Few-shot” prompting

Principle 2: Give the Model time to “Think”

• Tactic 1: Specify the steps to complete a task
• Tactic 2: Instruct the model to work out its own solution before rushing to a conclusion

Principle 1: Write Clear and Specific Instructions

• Express what you want by providing instructions as clear and specific as possible.
  → This helps to guide the model to desired output, and reduce the chance of getting irrelevant/incorrect responses.
• Clear ≠ Short
  → Prompts with longer instructions provides the clarity and context, so it brings much more precise and detailed output.

Tactic 1: Use Delimiters

• Utilize delimiters to clearly indicate distinct parts of the input, like specifying which part needs to be summarized.
• Delimiters can be anything like:

Triple quotes: """
Triple backticks: ```,
Triple dashes: ---,
Angel brackets: <>,
XML tags: <tag> </tag>

• Example:

✏️ What is the advantage of using delimiters?

• Avoiding Prompt Injection:
  Users are allowed to add input to our prompts. This behavior may contradict with the instructions to the models, so users get a unexpected or bad response.
  → In other words, delimiters can help to define and distinguish different regions so that hackers or people with bad intentions cannot revise the instructions that the model should follow.
• Example:

Tactic 2: Ask for Structured Output

With structured output, it assists us to parsing the model output easier and faster. In others words, you don’t have to process the output, such as eliminating undesired words/signs.

• Common output format: HTML, JSON, XML etc.
• Example:

Tactic 3: Ask the Model to Check whether conditions are satisfied

Check the assumptions required to do the task
→ Through making assumptions or setting edge cases, it teaches models how to address with the unexpected results and avoid from making mistakes.

Example:

• Pass assumption: using if/else statement to address different use cases.

text_1 is a sequence of instructions, so it enters the first “if” and gets processed to the format.

• Fail:

text_2 is not a sequence of instructions, so it can’t be rewrite into the format; it goes to the second “if”

Tactic 4: Few-shot Prompting

Few-shot prompting: a technique where the model is given a small number of examples, typically between two and five, in order to quickly adapt to new examples of previously seen objects.

Give successful examples of completing tasks. Then ask model to perform the task

Principle 2: Give the Model time to Think

If a model is given a task that is too complex, it may not be able to complete it within a short period or provide a short response, leading to potentially fabricated or incorrect answers.

✏️How to fix ?
1. Chaining: requiring a series of reasoning processes. (e.g. Step1, Step2, … Step N)
2. More time to respond: this implies allocating additional computational resources/costs)

Tactic 1: Specify the steps to complete a task

Chaining: Step1, Step2, …, StepN
Below are examples that perform multiple tasks with two different kinds of output.

• Example 1: Text output (cons: harder to parse)

In prompt_1, use step-by-step actions to break a complex task into smaller tasks.

The text output of the model. (hard to parse the data)

• Example 2: XML output (easier to parse)

Telling the model to use XML to format the output, which is better for subsequent processing.

Tactic 2: Instruct the model to work out its own solution before rushing to a conclusion

aka “More time to respond”

Example: Ask a model to do the math with a complex description.

• Wrong 🙅:

• Correct 👍:

Simplify the description and make model have enough time to respond

Result

Model Limitations

Hallucinations

• Model doesn’t understand how much knowledge it knows
  → Make statements that sound plausible but are not true

✏️ How to reduce hallucinations?

1. Find relevant information
   → Model should find relevant files or information firstly
2. Answer the question based on the relevant information

GitHub: MarsWangyang (Mars Wang) (github.com)

LinkedIn: Meng-Yang (Mars) Wang | LinkedIn