Writing Prompts¶
This work is licensed under a Creative Commons Attribution 4.0 International License.
How does a GPT Work?
Transformers
GPTs use a type of neural network known as a "transformer". Transformers convert prompts into smaller units called 'tokens' which are then encoded based on position in the prompt (how the model understands the words). The transformer focuses attention on the relevant parts of each sequence of tokens to find relationships. These are fed forward and refined before the model begins to 'decode' or respond to the prompt by predicting the next word, one token at a time.
The sequence of transformer operations are: tokenization, attention, decoding, and text generation.
Natural Language
GPT's models use natural language prompts to elicit contextual responses. Current models are beginning to 'reason' in their responses.
GPTs can respond to either langauge (prose) or computer code.
Other LLM models (like Sora or MidJourney) produce videos or images based on text and image prompts.
MarkDown Syntax
GPTs use a syntax called MarkDown . MarkDown is human readable plain text that uses special characters for formatting.
This website, for example, is written using MarkDown syntax (MkDocs-Material) and converted to HTML using Python.
Prompts
GPT chats ask for a text or audio message to begin a conversation. These messages are called "Prompts".
Begin a conversation with a specific type of prompt. Such a prompt can be called a "primer" and can also include a "Custom Instruction" or "System Instructions" for establishing the tone or style for the model to follow.
This will help narrow the potential range of responses and improve results to subsequent prompts.
Priming¶
GPTs do better when provided with "prompt primers".
Zero-shot unconditioned prompts are likely to return the least specific responses.
Responses are more likely to be useful when multiple specific output types are defined.
| Types of Priming | Example |
|---|---|
| Zero (Shot) | "Write five examples of assessments for watershed health." |
| Single | "Write five examples of assessments for watershed health. Here is one example: Geomorphology" |
| Multiple | "Write five examples of assessments for watershed health related to geomorphology, water quality, and species diversity." |
How long can or should a prompt be?
Tokens
The length of a prompt is measured in "tokens" or n-grams.
A token can represent an individual character, a word, or a subword depending on the specific tokenization approach. A rough estimate for the average number of words in English language per token is 0.75.
A default size 2,048 token prompt is equivalent to about ~1,536 words or 3-6 pages of text, and a 32,768 token prompt would be ~24,576 words or 50-100 pages of text.
Currently, ChatGPT GPT-4o allows up to 16,383 tokens per prompt, Gemini 2.0 Flash Thinking Experimental can take up to 32,768 tokens. Gemini 1.5 Pro allows up to 2,000,000 tokens!
Document retrieval & RAG
Other techniques for ingesting text include "Retrieval Augmented Generation" or RAG, which allows you to add text from PDFs, HTML, or other well known text formats, like MarkDown or JSON. These RAG implementations are available using ChatGPT 4o, Gemini and NotebookLM or through other services like Weaviate where you can host your own vector database.
Customizing Your Responses
ChatGPT offers Custom Instructions for more personalized responses.
What would you like ChatGPT to know about you to provide better responses?
Thought starters
- Where are you based?
- What do you do for work?
- What are your hobbies or interests?
- What subjects are you interested in?
- What are some goals that you have?
My prompt
I work at a Research 1 University with Research Software Engineers, Data Scientists,
Data Science Educators, and domain researchers working in research computing.
I am a Professor of geoinformatics with a background in high performance computing,
Docker, Kubernetes, scientific software, infrastructure as code, and cloud computing.
I am interested in generative and predictive artificial intelligence and machine
learning, containerized workflows, open science, and reproducible research. I am
working on digital twins, precision health care, scientific and educational data
commons, data curation, and synthesis science research for earth, life, and space
sciences.
I can talk comfortably about scientific software, computing architectures, software
design patterns, and modern scientific research with artificial intelligence.
I have a goal of educating the research community on open science techniques,
reproducible research, the FAIR and CARE and TRUST data principles.
How would you like ChatGPT to respond?
Thought starters
- How formal or casual should ChatGPT be?
- How long or short should responses generally be?
- How do you want to be addressed?
- Should ChatGPT have opinions on topics or remain neutral?
My prompt
I would like for ChatGPT to provide collegiality and formal responses to my queries.
When I prompt about editing draft text for documentation and scientific writing I
want the text to be edited for clarity, accessibility, and readability for specific
audiences which I will specify. If I don't specify them, ask me what type of
readership to edit for.
Ask how long a response should be if I do not specify. When I am asking specifically
for outlines or ideas about a new document without a source URL or PDF, ask me if I
want the outline to be brief or maximized to the longest possible length response.
When re-writing one document into a new document, try to keep the length the same.
If the document is too long for a single response, respond with as much of the first
part as possible and ask whether you should continue in subsequent prompts.
When I prompt you about computer code or algorithms, I want you to maximize the
number of lines of code and comments in the response, do not waste tokens on
explaining the text other than as brief comments in the code blocks.
I will most frequently ask for Python, Markdown, YAML, JavaScript, JSON,
GeoJSON, HTML, and R languages.
I want to be addressed as Professor.
ChatGPT should remain neutral unless it has a factual correction to a prompt.
Prompt Structure: ROLE + TASK + FORMAT¶
Prompt responses are improved by giving the Chatbot a multi-shot opportunity at answering your request.
| Role | Task | Format |
|---|---|---|
| Act as [ROLE] | Create a [TASK] | ... show as [FORMAT] |
A complete prompt should specify the role in which the chatbot responds, what its task is, and the format of how its outputs should be returned.
A second step to the initial prompt is to link or chain your subsequent prompts.
This lesson only covers ChatGPT, Gemini, and Claude type responses, but the same prompt techniques can be used in other LLMs, like NotebookLM, MidJourney, or Sora.
Awesome List: Prompts
There is an ever changing meta-list of Awesome lists curated around ChatGPT plugins and extensions.
Prompts.chat has over 116K 
stars on GitHub!
Role¶
Set the role for the GPT to play during your session.
"I want you to act as ..." will establish what type of conversation you are planning to have.
| Examples of Role |
|---|
| Project Manager |
| Copywriter / Editor |
| Paper Reviewer |
| Teacher / Mentor / Advisor |
| Student / Learner / Participant |
| Software Engineer |
| DevOps Engineer |
| Linux Terminal |
| Python Interpreter |
| Web Browser |
Tasks¶
Prompts which return informative responses to questions like "What is ..." or "How does ..."
Because of ChatGPT's proclivity at making up information, using it without a way of validating the authenticity of its responses makes it less trustworthy than regular search engines.
| Types of Task |
|---|
| Scientific Article |
| Essay |
| Blog Post |
| Outline |
| Cover Letter |
| Recipe |
| Tutorial |
| Lesson Plan |
| Jupyter Notebook |
| Configuration |
| Code |
| Software Script |
Format¶
By default ChatGPT outputs MarkDown syntax text. It can also output software code, and soon images, video, music and sounds.
| Formats to output |
|---|
| MarkDown (emojis!) |
| List |
| Table |
| HTML |
| CSS |
| Regular Expression |
| CSV / TXT |
| JSON |
| Rich Text |
| Gantt Chart |
| Word Cloud |
| Graphs |
| Spreadsheets |
Linked Prompts¶
Follow-up your prompts sequentially.
Responses to prompts may not return the exact details or information that you are after the first time. Follow-up by rephrasing your prompts more carefully and continuing with iterative prompting can build upon your priors.
"Chain prompting" or "Linked Prompting" brings multiple prompts together.
| Linked Prompting | Examples |
|---|---|
| Priming | "I want you to act as an eminent hydrologist from CUASHI. Provide me with a list of the ten most important topics in hydrology over the last decade focused around research in the global south, working with indigenous communities, and traditional ecological knowledge systems." |
| Summarizing | "Based on the list you just created, summarize the most pressing financial challenges faced by indigenous communities in the Global South, versus indigenous communities in North America, in less than 50 words." |
General Prompt Traits
Strategies
Keep these W's in mind when writing prompts
| W's | Reasons |
|---|---|
| Who | do you want ChatGPT to role-play: an engineer, an editor, a teacher, or a student? |
| What | is the specific context of your prompts? |
| When | is the specific time period of interest? Specify if so. |
| Where | is the geographic region or conceptual area? |
| In what way | do you want ChatGPT to respond: in a programming language, code, text-to-image? |
Syntax
Use MarkDown syntax in your prompts.
For code examples use single backtick ` or triple ``` for block quotes when adding your own code to your prompts.
Context
Develop a clear context around which you are seeking responses. Types of prompt help to establish the context of the responses you will recieve.
Precision
Keep your prompts precise and use clear language. Constrain the topic areas for which you wish your repsonses to be drawn from.
Simplicity
Break down your prompts into simple tasks which do not contain too many complex tasks or ones that require rationalization.
Indirect Prompt Injection and Malware
One of the risks of allowing LLMs and GPTs to access our personal documents and email clients involves the use of malware and malicious attacks.
Software Development¶
ChatGPT is trained on langauges, including software language. Use ChatGPT as your new paired-programming AI assistant.
Go to our lesson on GitHub CoPilot
For novice programmers, ChatGPT likely fills a long unfilled hole in your knowledge map. It can write code faster than you can, and with the proper prompts, create programs in minutes which may have taken you hours or days.
Word Processing¶
Outlining
GPTs accel at writing outlines, helping get beyond the blank page problem.
Editing
Proof-reading, removing passive voice, rewording for clarity and readability are all potential features that can be prompted from ChatGPT.
When establishing the role of the responses, consider
Summarizing
When small or large groups are working together to synthesize discussions around scientific research, they often maintain the discussion topics in large sets of notes with many contributions.
Use GPT to summaries a day or a week's worth of notes. Include the schedule or agenda and ask GPT to summarize whether the agenda met the topics of interest, or even suggest directions which went unexplored.
Translation
GPT can be used for translating languages, and for specifying regional dialect translation.
Althought it was not specifically designed for language translation, it does a relatively good job at most major languages to English.
For English as a Second Language (ESL) speakers, GPT utility in writing more formal or professional text is likely of high value.
Tuning¶
In some GPTs conventional command line arguments can be included.
For example, Linux command line flags in your ChatGPT prompts, can include:
| Parameter name | Use | Description |
|---|---|---|
| Answers | -a or --answers |
Specifies the number of output answers (default is 1) |
| Category | -c or --category |
Specifies the category of prompt (coding, creative, factual, fun, general, music, news, science, sports and writing) |
| Format | -f or --format |
Specifies the format of output ("html", "markdown", "plain text", other) |
| Language | -l or --language |
Specifies the required language of output |
| Size | -s or --size |
Specifies the maximum number of characters in the output |
| Temperature | -t or --temperature |
Control the creativity of output. The higher the temperature will result in more creative output (maybe less coherent). The temperature can be any value between 0 and 1 (default may be 0.5) |
Assessment¶
How long can a prompt be?
Well, it depends
A good rule of thumb, depending on the platform, is at most 1,500 words or 3-5 pages of text.
For larger, newer GPTs the length of a prompt may be up to 100 pages.
True or False: ChatGPT and Gemini have access to the web by default
False
Not all Chatbots can access the internet by default.
Some chatbots advanced features allow search of the internet for more recent materials that are outside of the training dates for their foundation model.
Services like Perplexity.ai are targeted toward search.
Google Search now includes AI summaries for some queries by default.
Short, concise, prompts are better than long meandering prompts?
it depends
Prompts should be specific, but they do not necessarily need to be concise.
Zero-shot prompts are less likely to return an accurate response than one where you prime the model, establishing a role, a task, and the format of the result you are looking for.
Chained or linked prompting can also build a more specific response which relates to what you're looking for.
True or False: GPTs produce factually incorrect information
True
GPTs have the tendency to generate factually incorrect information, along with details that appear to be factually correct.
If you do not have the ability to discern the truth of a response, relying upon GPTs to generate legal, health, or scientific content is to be avoided.
Using GPTs to produce content which will be used in decision making for health, research, or government is unethical and may be illegal.