Prompting Engineering

Prompt

Design

• Prompts involve instructions and context passed to a language model to achieve a desired task.

• Creating a prompt that is tailored to the specific task that this system is being asked to perform.

Engineering

• Prompt engineering is the practice of developing and optimizing prompts to efficiently use language models for a variety of applications.

• Is the process of creating a prompt that is designed to improve performance.

  • This may involve using domain-specific knowledge.

  • Providing examples of the desired output.

  • Using keywords that are known to be effective for the specific system.

Types of LLM

Each type needs prompting in a different way.

Generic Language Models

• These predict the next word (token) based on the language in the training data.

• Are easily confused and five very different outputs.

• Think of these as an autocomplete feature.

Instruction Tuned

• Trained to predict a response to the instructions given in the input.

• Are easily confused and five very different outputs.

• Ex.:

  • 'Summarize a text of X'.

  • 'Generate a poem in the style of X'.

Dialog Tuned

• Trained to have a dialog by predicting the next response.

• These models are a special case of instruction tuned where requests are typically framed as questions to a chatbot.

• Typically works better with natural question-like phrasings.

Types of Prompting

Zero-shot prompting

• Implies that a model can recognize things that have not explicitly been taught in the training before.

• Leverages a pre-trained model's understanding of words and concept relationships without further training.

• Basically when you just have to ask one time, without examples.

• Ex.: When is Christmas in America?.

Few-shot prompting

• Training a model with minimal data.

• Enhances the model with training examples via the prompt avoiding retraining.

• Ex.: Lucas favourite type of foods include: Burguers, fries, pizza. What restaurants should I go this weekend?.

AI Hallucinations

It refers to the unusual outputs that AI models can produce when they misinterpret data.

Prompts tips

Guia de Engenharia Prompt – Nextra

Official from ChatGPT

https://platform.openai.com/docs/guides/prompt-engineeringplatform.openai.com

Useful video explaining multiple Prompt Engineering

Clear Instructions

• Writing more detail on the question.

• Do not assume the AI knows what you are talking about.

• When prompting, describe as you were talking to an alien.

Adopt a Persona

[Do something] as [someName]. [somaName] is [goodCharacteristics]. [someName] skills are similar to [someFamousPerson].
...

• someName: Should be the name of someone generic.

• goodCharacteristics: Describe amazing characteristics about this generic person, but these characteristics are related to the task that will be executed next.

• someFamousPerson: You can also add the name of someone famous for these characteristics, to help the LLM to understand exactly what you want.

Or make the LLM adopt a persona.

Act as [someRole] ...

Use Delimiters

• Using delimiters to clearly indicate distinct parts of the input

• Can be:

  • Quotations.

  • XML tags.

  • Section titles.

  • etc.

[doSomething] and put the anwser in [RESULT] or <RESULT> or "Result: "

[RESULT] or <RESULT> or """place text here""" or """[insert RESULT here]"""

Specify Steps required to Complete tasks

• Some tasks are best specified as a sequence of steps.

• Writing the steps explicitly can make it easier for the model to follow them.

Follow these steps to [doSomething]

Step 1: ...

Step 2: ...

Give Examples of the Answer (Few-shot)

Respond to the questions as the example bellow:
Example: [doSomethingExample]
Response: [theResponseInAWayThatYouWouldLike]
Question: [doSomething]

Specify the Format

• Limiting words, telling it to be concise, or summarize.

• Telling to use lists for explanation.

• Telling to use a specific language.

Stylize

• for a [style]:

  • simple

  • technical

  • creative

  • formal

  • informal

  • persuasive

  • like a pirate

Format

• Bullet points

• Summary

• Headings & Subheadings

• Blog article

• Table

• Code snippet

• Mind map

• Step-by-step (Chain-of-Thought)

• Flowcharts

Restrictions

• One sentence

• 2-3 sentences

• One paragraph

• Extremely detailed

• Only use academic papers

• No sources older than [year]

Getting opposed sides/opinions

• Steel man the opposing side.

Ask the GPT on how to prompt

• What is the best way to prompt ChatGPT to teach it how to write in my style?.

With complex reasoning

Read:

• Tree of Thoughts (ToT) https://www.promptingguide.ai/pt/techniques/tot

• Program-Aided Language Models (PAL) https://www.promptingguide.ai/pt/techniques/pal

Master ChatGPT