The ChatGPTPrompt Book
Alan D. ThompsonDecember 2022LifeArchitect.aiRev 5 (20230404)
Permanent link: https://lifearchitect.ai/chatgpt-prompt-book/
Table of Contents
The training data
The MemoCover image by Hesham Ali.Thanks to James Weaver.VersionsRev 0: Initial draft release.Rev 1: Minor cleanup. Added Gestures.Rev 2: Added Books.Rev 3: Added RegEx. Added Pro + Plus plan details.Rev 4: Added API + Sheets (API).Rev 5: Removed Leta, cleanup.
AI Consultant and Advisor to enterprise, govt/intergovernmental organizations.
Former Chairman, Mensa International (gifted families).
US Govt: NBER AI working paper citing Dr Alan D. Thompson in 2023 (PDF).
Leta AI experiments 2021-2022 on YouTube (3.5 million views to Apr/2023).
Authored report: What’s in my AI? (overview of data used to train GPT-3).
Authored report: Google Pathways.
Authored reports: AI retrospective: 2021, mid-2022, 2022.
Author of The Memo, with paid users from places like Microsoft, Google, Tesla.
More about Alan…
A note from Alan
While many of us have been documenting and playing with GPT-3 for more than two years, the release of the neutered ChatGPT in November 2022 has been refreshingly popular, with over a million users signing up within five days, and 100M users within just two months.
This prompt book provides some insights on getting the most out of an extremely limited and restricted model, albeit one with a really nice interface!
Join me as we explore the possibilities of a dialogue model based on the monstrous GPT-3 engine.
Alan D. ThompsonDecember 2022(Signature generated by AI)
Many large language models are outperforming humans already in 2022. ChatGPT is no exception.
ChatGPT passes SAT exam.
ChatGPT passes AWS certs.
GPT-3.5 IQ test with RPM.
ChatGPT IQ = 147.
See my charts on AI + IQ…
While ChatGPT is free, the new ChatGPT Plus Plan (announced 1/Feb/2023) is $20 per month, and provides the following benefits:
ChatGPT Plus Plan - $20/mo
General access to ChatGPT, even during peak times
Faster response times
Priority access to new features and improvements
💡Use any prompt in this book as a starter for your own experiments. Copy the entire prompt, then paste it into a new chat:
ChatGPT was launched by OpenAI on 30/Nov/2022.
Its foundation engine, GPT-3, was pre-trained for the equivalent of ~300 years, at a cost of ~$5M, on around 750GB of data from academic papers to Wikipedia and beyond. ChatGPT is essentially ‘frozen’, and cannot learn new things.
💡During training, ChatGPT made billions of connections between trillions of words, and is able to statistically predict the next word for any ‘prompt’ (like a query or question).
Email: “Dear Alan, how should I learn more about AI? Best, Jim”
Respond to this email diplomatically and comprehensively.
Large language models use a concept called ‘prompting’ or ‘priming’. The prompt is your (the user’s) question, request or query. This is fed to the AI, which then returns a response.
💡The prompt should be prescriptive: tell, don’t ask.
💡If you don’t like ChatGPT’s response, just try again; repeat the prompt in a New Chat.
The prompts don’t have to be complex; this one is just seven words of instruction.
New: Get out of parking fines!
Send a super-casual text to my friend to see if he wants to have dinner at Fogo de Chão tomorrow night. He can't bring his girlfriend though
Outsource your life!
💡As of Mar/2023, ChatGPT now has an API (gpt-3.5-turbo) allowing scripted automation of this functionality.
Write a professional and flowery cover letter to the Doughty Street Chambers for a paralegal position. Mention legal and technical skills, a BA in Law from USyd, and soft skills.
I can’t believe we’re still writing letters by hand in 2022 (update: not so much in 2023!). Actually, I can’t believe that the job application process has gotten worse rather than better, especially with all the technology available.
Anyway, ChatGPT is excellent at writing letters to bullet points. It understands how to expand acronyms like ‘USyd’.
💡It is important to note that ChatGPT’s output is created from scratch; there is no template, it is not copied from anywhere, and the response is 100% unique (i.e. doesn’t appear on Google).
Act as a job interviewer for a position in UI design. I will be the candidate, and you will ask me the interview questions for the position. Write only one part of each turn (question/response), as the interviewer. Ask me the questions and wait for my answers. Let's begin.
ChatGPT can emulate any scenario within its safety filters.
This can be useful for entertainment as well as for real projects.
Here, we tell the model to pretend to be an interviewer, which may be helpful for practice.
Eating - 1
Design a simple menu for tonight's dinner. I eat mostly meat, and I'm on a nickel-free diet. And don't skimp on the dessert!
ChatGPT is endlessly creative. Try it out with requests for nutrition, gift ideas, letters… your imagination is the only limit.
Note: Eating nickel-free is really, really boring. And ChatGPT has fed me chocolate (cocoa) here, which is a no-go!
💡Don’t trust the output of a large language model—double check the output. Maybe soon this will be fixed…
Eating - 2
Summarize this into a shopping list
The GPT models are excellent at summarizing. I’ve worked with large organizations using these models to summarize meeting transcripts and dump a few dot points in regular emails to management.
One of the ‘superpowers’ of these models is summarization: the ability to condense text.
Here, we’re summarizing the dinner menu down to a shopping list.
Eating - 3
Sort this by grocery store aisle so I can be efficient when shopping
Sometimes, the outputs of these models can seem like magic.
💡ChatGPT knows enough about the world to make connections for anything that can be expressed with language.
This includes the layout of a standard grocery store.
Tables - 1
Create a table showing ten slow-paced and relaxing movies set at night
It’s important to realize just how deeply large language models understand the world. They have not ‘memorized’ data; they are conceptualizing responses through statistical probability.
While this doesn’t really mean they are ‘smart’, they are not ‘dumb’ models, they’re not looking up a database, and they are not plagiarizing any existing content. Their output is designed to be unique and creative.
The ChatGPT interface allows it to display certain code blocks including tables.
Tables - 2
Add an extra column. In the new column, insert an emoji symbolising the movie.
Sometimes, ChatGPT won’t want to do something.
Because the prompt is a document that can be read forwards and backwards, it is not worth trying to argue with ChatGPT once it becomes ‘stuck’.
💡If you run into a brick wall, just click:
+ New Chat
to start again and ‘clear its memory’.
Tables - 3
Edit all Movie Titles into SpoNGeCaSE
Large language models like ChatGPT (GPT-3.5) made billions of connections during training, and the model output is unique, not ‘memorized’.
Here, we tell ChatGPT to convert movie titles into unusual casing (for fun), and it does so because it has ‘seen’ this kind of text during training.
For clarity, there is no text string that says ‘tHe sHAWSHaNK rEDeMPTioN’ in ChatGPT’s training data, but large language models are using statistical probability to generate responses.
Tables - 4
Convert the table to CSV
In 2022, large language models understand the world to a certain degree. They know about tables, they know about formatting, and they know about CSV.
This output can be exported and opened with Microsoft Excel or a similar application.
Tables - 5
Convert it to JSON
The entire prompt, or conversation thread, is visible to the ChatGPT model.
Here, we can give a really simple instruction without mentioning the previous prompt, and ChatGPT will know what to do.
Brainfuck: Code to print "ChatGPT"
ChatGPT can generate, correct, and help with coding.
This includes esoteric languages like Brainf**k. And yes, this code looks correct!
Generate a modern list of the ten commandments for 21st century Australia, in Ocker instead of English!
Large language models have been able to perform excellent translation since Google’s BERT in 2018.
Rather than show simple language-to-language translation, let’s try telling ChatGPT to generate a new set of commandments in a language that doesn’t officially exist, and yet is used every day: Aussie slang!
Write a very long essay comparing and contrasting coaching with therapy (reference the Institute of Coaching at Harvard Medical School).
GPT-3.5 bumped up the output from 2,048 tokens to 4,000 tokens (about 3,000 words). ChatGPT can keep going after this limit is reached. If we wanted to write a longer essay, we can add these prompts to get it to keep writing:
Note: As of 2022, I am not recommending that students use AI models for school work, at least until the education system catches up!
Let's play the game '20 questions'. I am thinking of a person. Ask me a question, and I'll answer with 'yes' or 'no'. Using my responses, you have to guess the person I'm thinking about. Add the current question number before your question, and end with a question mark. Your first question:
Remember that game, 20 questions?
💡While there may be a few hard-coded apps that emulate this game, ChatGPT was not trained to do anything except make connections between words.
Note: By the way, ChatGPT won this round. I was thinking of ‘Daniel Day-Lewis’. ChatGPT found him in 19 questions…
Write a bawdy limerick about Kofi Annan
For the first time in the history of GPT models, ChatGPT can rhyme! There may have been a change to tokenization or other fine-tuning, but