Overview

Types

Completion Prompts

Suitable for single-turn tasks where the model generates a response based on a single input prompt
Conversation context/history and role seggregation is not essential
Works well for content generation, summarization, question-answering etc.

Chat Prompts

Designed for multi-turn conversations with multiple roles (user, assistant, system etc.)
- System prompt - indicates what task the AI is performing and what tone it should use
- User prompt - the prompt from user that the AI should reply to
Maintains conversation context (see context window) by processing the entire conversation history
Ideal for chatbot applications and tasks requiring back-and-forth interactions

Prompting Guidelines

Structure

Break down large prompts into steps
- Use delimiters for organizing the prompt
  Stop Sequences
  - Can be used to signal the end of an example in few-shot prompting
  - Helps to structure the input data for the model
  - Triple backticks, quotes, double hashtags etc.
- Break large prompts into "sub-prompts" if needed
- Context: Brief introduction or background information
- Include topic, industry/field, relevant links, keywords, datasets, text chunks etc.
- Constraints or restrictions
- Persona: Role or expertise/profession the prompt should use
- Use phrases such as Act as, Your role is to, Imagine you are etc.
- Add relevant details such as field of work, company name, time period, region etc.
- Goal/Intent: Define what the prompt should do
- Describe objective
  - Use action verbs such as Analyze, Generate, Simplify, Summarize etc.
- Mention any focus areas using phrases such as Focus on etc.
Useful Action Verbs and Adjectives
- Verbs
  - Analyze, Compare, Convert, Customize, Describe, Evaluate, Explain, Generate, Improve, Optimize, Organize, Proofread, Provide, Revise, Rewrite, Share, Simplify, Translate, Write
- Adjectives
  - Clear, Concise, Conversational, Professional
Output
- Audience: Layman, x year old, type of professional, demographics (age/gender/location) etc.
- Tone: Formal, Business, Casual, Academic, Humorous etc.
  - As a Famous Personality e.g. Shakespeare, Tagore etc.
- Other Specs:
  - Specify constraints (Number of words, points, lines, paragraphs etc.):
    - Summarize using x bullet points, numbered lists etc.
    - Make your response as short as possible.
  - Specify length of output
    - Use at most x words.
    - Use no more than x sentences.
    - Give your answer as a single word, either "this" or "that".
  - Ask for additional details
    - examples and/or analogies
    - for and against arguments
  - Specify what not to do
    - Do not include x, y and z
Clarify: Ask the prompt to ask clarification questions if needed
- Ask me questions (one at a time) to (add relevant text based on the ask)
Validate: Ask for proof for reducing hallucinations and fabricated responses
- Ask model to find relevant information, then answer the question based on the relevant information
- Provide relevant sources and citations

Code chunks in Prompts

Include code chunks in a prompt by enclosing them in single backticks followed by the name of the language:

Some text
`python
some python code
`
More text

Tactics¹

Tactic 1: Use delimiters

Triple Quotes: """
Triple backticks: ```
Triple dashes: ---
Angle brackets: <>
XML tags: \<tag>\</tag>

Sample

text = f"""
A paragraph of text
"""

prompt = f"""
Summarize the text delimited by triple backticks into a single sentence.
```{text}``` 
"""

Tactic 2: Ask for a structured output in a specified format

JSON, HTML etc.
Markdown with structured sections and headings

Sample

# Example 1a - Simple JSON with specified keys
prompt = f"""
Generate a list of three bollywood movies along with their director and year of release. 

Provide them in JSON format with the following keys: 
Movie Name, Year of Release, Director.
"""

# Example 1b - Simple JSON with example
prompt = f"""
Generate a list of three bollywood movies along with their director and year of release. 

Provide them in JSON format as in this example:

{
    "Movie Name": "Sholay",
    "Year of Release": 1975,
    "Director": "Ramesh Sippy"
}
"""

# Example 1c - Structured JSON
prompt = f"""
Your task is to perform the following actions: 
1 - Write a parable.
2 - Summarize it in a single line in Bangla.
3 - State the moral of the parable.
4 - Output a json object that contains the following keys: title, moral.

Use the following format:
Parable: <text to summarize>
Summary: <summary>
Moral: <moral of story>
Output JSON: <json with title and moral>
"""

# Example 2 - HTML Table with specified columns and datatype
prompt = f"""
Your task is to perform the following actions: 
1. Prepare a list of 10 bollywood movies
2. Find out year each of the movies was released
3. Find out if the movie was a hit or a flop at the box office.

Output the response as a table having three columns.
In the first column include the name of the movie.  Use "Name" as the column title.
In the second column include the year of release (Use "Year" as the column title ) and in the third column, include the box office result. Use "Hit or Flop" as the column title and format the value as a boolean.

Give the table the title 'Movie Stats'.

Format everything as HTML that can be used in a website. 
Place the description in a <div> element.

Output in markdown format as well. 
"""

# Example 3 - Markdown
prompt = f"""
Respond to the following in markdown format. Do not wrap the markdown in a code block - respond just with the markdown: 
```{text}```
"""

Tactic 3: Ask the model to check whether conditions are satisfied

Sample

text = f"""
x=1 y=2 x+y=3
"""

prompt = f"""
You will be provided with text delimited by triple quotes. 
If it contains mathematical equations, re-write the equation as a list of steps in the following format:

Step 1 - ...
Step 2 - …
…
Step N - …

If the text does not contain contains a mathematical equation, then simply write \"No equation provided.\"

\"\"\"{text}\"\"\"
"""

Tactic 4 - Few-Shot Prompting

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

Sample

prompt = f"""
Your task is to answer in a consistent style.
<question>: Sholay.
<answer>: The movie "Sholay" was released in the 70s and started Amitabh and Dharmendra in the lead roles.
<question>: Karz.
<answer>: The movie "Karz" was released in the 80s and started Rishi Kapoor and Tina Munim in the lead roles.
<question>: Mahaan.
"""

Tactic 5: Specify the steps required to complete a task

Sample

prompt = f"""
Perform the following actions: 
1 - Write a parable.
2 - Summarize it in a single line in Hindi.
3 - State the moral of the parable.
4 - Output a json object that contains the following keys: title, moral.

Separate your answers with line breaks.
"""

Tactic 6: Instruct the model to work out its own solution before rushing to a conclusion

Sample

prompt = f"""
Your task is to determine if the student's solution is correct or not.
To solve the problem do the following:
- First, work out your own solution to the problem. 
- Then compare your solution to the student's solution and evaluate if the student's solution is correct or not. Don't decide if the student's solution is correct until you have done the problem yourself.

Use the following format:
Question:
- question here
Student's solution:
- student's solution here
Actual solution:
- steps to work out the solution and your solution here
Is the student's solution the same as actual solution just calculated:
- yes or no
Student grade:
- correct or incorrect

Question:
I'm producing a television show and need help
working out the cost associated to the payments for the actors. 
- Lead actor charges $100 / episode
- Lead actress charges $90 / episode
- I negotiated a contract for character actors that will cost me a flat $1K for the first 10 character actors for the entire show, and an additional $25 per additional character actor / episode
What is the total amount I need to spend on the actors as a function of the number of episodes. Assume that I end up using a total of 12 character actors per episode.

Student's solution:
Let x be the number of episodes.
Costs:
1. Lead actor charges: 100x
2. Lead actress charges: 90x
3. Character actor charges: 1,000 + 300x
Total cost: 100x + 90x + 1,000 + 300x = 490x + 1,000

Actual solution:
"""

Basic Examples

Provide context
- Here is some information to include in /for reference
Summarize
- Web Pages: Paste the link of the web page followed by the word summarize
- Write a summary of the book xyz by abc
- Provide an overview of some event
Compare
- Compare and contrast the concepts of (or the words ..) a and b
Explain/Describe
- Explain the concept of xyz in simple terms
- Explain the concept of xyz using abc analogy
- Describe the top x abc
Generate
- Write a simple recipe for xyz
- Share x tips for abc
- Help me brainstorm xyz
Translate/Convert
- Translate the following from one language to another language
- Convert the following to xyz format
Proofread/Review
- Proofread and correct any errors in the following
- Rephrase/Structure the above for maximum impact
- Review the above and suggest areas for improvement in terms of tone and engagement
- Is there something I should add to make this resonate better?
Follow up Instructions: Additional prompts to make changes to initial output
- Provide x alternate versions for
- Make the content more/less xyz
- Incorporate xyz to make the output more abc
- Add more details / expand on xyz
- Focus on xyz

See Prompt Samples for more examples

Prompt Caching

OpenAI

Cache hits are only possible for exact prefix matches within a prompt
- To realize caching benefits
  - Place static content like instructions and examples at the beginning of your prompt
  - Put variable content, such as user-specific information, at the end
  - Applies to images and tools as well
Cached input is 4X cheaper
Pricing

Anthropic

Have to tell Claude what you are caching
Price is 25% MORE to "prime" the cache
- Price is 10X less to reuse from the cache with inputs
Pricing

Gemini

Supports both 'implicit' and 'explicit' prompt caching
- Implicit caching
  - Automatically enabled on most Gemini models
  - No cost saving guarantee
- Explicit caching
  - Can be manually enabled on most models
  - Cost saving guarantee
To increase the chance of an implicit cache hit:
- Try putting large and common contents at the beginning of your prompt
- Try to send requests with similar prefix in a short amount of time
Pricing

Based on DeepLearning.ai ↩

Overview

Types

Completion Prompts

Chat Prompts

Prompting Guidelines

Structure

Tactics1

Basic Examples

Prompt Caching

OpenAI

Anthropic

Gemini

Tactics¹