Blog: ChatGPT Prompt Engineering for Developers

Posted on Tue 07 May 2024 in blogs

Two Types of LLM

  1. Base LLM:

    • Predicts next word, based on text training data.

  2. Instruction Tuned LLM

    • Fine-tune on instructions.
    • RLHF: Reinforcement Learning with Human Feedback.

Guidelines

  • Principle 1: Write clear and specific instructions.

  • Tactic 1: Use Delimiters.

  • Tactic 2: Ask for structured outputs.
  • Tactic 3: Ask model to check if conditions are satisfied.

  • Tactic 4: Few-shot prompting. Give successful examples then ask to perform task.

  • Principle 2: Give the model time to think.

  • Tactic 1: Specify the steps required to complete a task.

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

  • Limitations

  • Hallucinations: These are fabricated ideas which is made up response with no relevance to real life scenarios.

Iterative Prompt Development

  • Idea -> Implement -> Experimental Result -> Error Analysis.
  • Limit the number of words/sentences/characters.
Your task is to help a marketing team create a description for 
sale season of a retail stores of the products based on 
technical fact sheet.

Write a product description based on information provided
in the technical specifications delimited by two backtics.

Use at most 50 words.

Technical specifications: ``{{text}}``
  • Ask to focus on specific details.
  • Ask to organize in table.

Summarizing

  • You can summarize the product for specific relevant area.
  • Limit the number of character/words/sentences.
  • You can also extract the information rather than summarize.

Inferring

  • Infer sentiments from text.
  • Example
What is the sentiment of the following tweet, which is delimited with double backticks?

Tweet: ``<tweet>``
  • You can also get answer in one word like positive or negative using text like,
...
Give your answer as a single word, either "psoitive" or "negative".
  • Identiy the types of emotions,
...
Give me a list of emotions that writer of the following tweet is expressing. Include no more than
six items in the list. Format your answer as lower-case separated by commas.
  • Extract information,
Identity the following item from tweet:
- User the tweet is refering to.
- Product the user if tweeting about.

Format your response as JSON object with 
referred_user & product as keys. If information is not
present then use "unknown".
  • You can also merge the above under one task.
  • Infer topic from text.
Determine five topics that are being discussed in the following text,
which is delimited by two backticks.

Make each item two or three words long.

Transforming

  • It can perform following task but not limited to these:
  • Translation

text Translate the following English text to German:

text Tell me what language is this:

text Translate the following text to French in both formal and informal forms:

  • Spelling/Grammer Checking

text Proofread and correct the following text and rewrite the corrected version. If you don't find the error then just say "Everything is Alright".

  • Format conversion

text Translate the following JSON into HTML format:

  • Tone adjustment

text Translate the following slang to business letter:

Expanding

  • You can ask you model to ask as assistant and reply to the customer with required setiments.
  • Make sure you use it responsibly and it generates relevant response.

Chatbot

  • We can set various role based on requirements. Allowed values are:
  • system: You can set the behaviour of the model.
  • assistant: This is what model outputs.
  • user: This is what useer inputs to the model.
  • Each conversation is seperate context.

ml