Boost Your AI Skills: Learn from the Best

Learning these, and other prompt techniques, is important because:

• Backed up by scientific papers — the prompt techniques discussed in this article potentially increase performance by 295% (Liam Ottley, 2024).

This tutorial's prime function is to explore prompt techniques.

We are empowering you with prompt understanding and knowledge. This helps you avoid reliance on rigid templates and paid tools.

The Prompt Structure

• Role
• Task
• Specifics
• Context
• Examples
• Notes

The format | structure shown here is used in prompting AI Systems.

It is a proven technique adopted in Liam Ottley’s AI agency, Morningside  . These systems require LLM outputs that are accurate and error-free. Prompts for AI systems are efficient and effective on cheaper LLM models.

In building this tutorial, I will start with a base 'non-engineered' immature prompt.

I then discuss each of the sections Role, Task, Specifics, Context, Examples, and Notes, and apply what we learn to the final prompt.

This is not the only successful prompt structure in use. It does, however, mirrors many similarities, simply because it adopts proven principles.

Use these structures as a map to understanding principles. Most prompt situations require a nuance. Knowledge of principle informs the appropriate nuance to apply.

In future articles, I will explore the AUTOMAT and CO-STAR frameworks put forward by Maximilian Vogel  .

Large language models (LLMs) can output brilliant and exact answers or solutions if — and this is a big if — “you specify your wish in a perfect prompt (Vogel, 2024).” This requires you to “craft your prompts masterfully and meticulously (Vogel, 2024).”

A ^*forMulA^* for Initial Customer Profile (ICP) Research

We adapt the approach used by Morningside   to a conversational prompt process for ICP. This is the secondary outcome, a *'fOrMuLA'* to follow when researching an Initial Customer Profile (ICP) using LLMs.

With this ICP research, we aim to find the hidden, the not-obvious.

Tackle the ICP discovery from an objective perspective. Not to do ICP research from our subjective perspective.

We will see later that the non-engineered prompt [almost] guarantees a subjective, mediocre result.

If you run this prompt, a probable outcome is for the LLM to ask generic questions related to your target audience. Questions such as:

1. What product or service does your business offer?
2. Is your product/service aimed at consumers (B2C) or other businesses (B2B)?
3. What problem does your product/service solve for your customers?
4. Can you describe your ideal customer in terms of demographics (age, gender, location, income level, education level, etc.)?

This is not specially helpful as you are giving it the answer. Especially for question 4, as this is the information we are researching.

I found the LLM output to be a regurgitation of my answers to the questions asked, with no additional value. I do not need an LLM to get this type of information — I already have it.

We see the above process leads to a subjective answer. We are doing research rooted in our perspective.

To find the non-obvious, we do objective research, rooted in the eyes of the customer.

Select a role that gives an advantage for the task, eg. a Marketing Professor for a marketing task. A YouTube Channel Manager for a YouTube task.

• Adding a role shows increased accuracy by 10.3% (Kong et al., 2023).

Perplexity gives the following summary definition of Role Prompting as a technique.

"In summary, role prompting is a prompt engineering technique that involves instructing the AI to adopt a particular persona or perspective in order to generate text that is tailored to the desired style, tone, and purpose. "

This helps “immerse the model in the role and can improve performance on downstream tasks (Liam Ottley, 2024).”

Give the Persona Additional Context

• Adding additional context to the role results in a total accuracy increase of 15 - 25% (Liam Ottley, 2024) (Kong et al., 2023).

A trick I discovered was to give the role context. For example.

You are doing research on MECLABS Perceived Value Differential heuristic. You want clarity on a certain point and turn to ChatGPT for contextual information.

A common marketing persona for this type of prompt is that of a Professor of Marketing at Harvard University. I find giving the Professor context improves the quality of output.

<persona>
You are a Professor of Marketing at Harvard University
</persona>
<persona_context>
You are analysing work submitted by an undergraduate. The subject of the work submitted is [enter subject matter]. Your contextual role is that of a mentor to the undergraduate, offering expert knowledge, insights, guidance, and advice.
</persona_context>

Ottley suggests a similar technique

“Enrich the role with additional words to highlight how good it is at that task.”
“Complimentary descriptions of their abilities further increased accuracy (~15-25% increase total!)” 

For example, act as a YouTube Channel Manager:

Our Prompt — Adding Role

Notes   : Here we added a role and embellished the role definition with skills, aptitudes, ethos and unique abilities.

Task definition is the one we are most familiar with. Without a Task, there is no prompt. A few points to keep in mind:

• Tasks start with a verb (generate, summarise, compile);
• Have a clear objective — know what outcome you want to achieve.”Even the smallest mistake or misconception in a natural language prompt can give wildly different results than expected (Delisi, 2024).”;
• Be specific — “give clear and complete instructions (Delisi, 2024).”;
• Avoid ambiguity — words such as ‘it’, ‘this’ or ‘that’ are ambiguous. Specify what you are referring to. “Summarise this text” is better stated as “Summarise the text delimited by the tags <text>”;
• Be simple and concise — “The prompt should be simple and concise to reduce the chances of ambiguity or confusion (R, 2023).”

This article by phData touches on two points worth mentioning.

1. Use the Golden Rule — “Only ask questions that you know the answer to (Delisi, 2024).” and
2. Knowledge Discovery is a Dangerous Task — “Remember that the LLM will be confident in giving an incorrect and plausible-sounding answer (Delisi, 2024).”

In our use case, ICP research, the objective is finding the unknown. We look for an unseen answer. We are performing knowledge discovery.

This brings home VERY VERY important points when using LLMs.

1. Claim ownership and responsibility of the output;
2. Retain the seat of your power — do not abdicate to the LLM; and
3. Always be discerning, observant and verifying.

The hypothesis presented by the LLM output is verified, either in your mind or in a test of some form.

By observing these three principles, ultimately in a constant state of verification, you are answering questions you know the answer to and removing the danger from knowledge discovery.

Chain of Thought and Take a Deep Breath

Considering the task component, there are methods to improve the output quality. I discuss two:

1. Think step-by-step (Chain of Thought); and
2. Take a deep breath.

With Chain of Thought (CoT), tell the model to ‘think step-by-step’ or give it a step-by-step process to follow.

Research results show that CoT gives a 10% and 90% increase in accuracy on simple and complex tasks, respectively.

We encourage the model not to rush to an answer. ‘Take a deep breath’, ‘Think carefully’, and ‘Use a scratch pad’ are techniques that encourage the LLM to 'take time to think'.

Delimiters and Variables

We use delimiters in our prompt, so a brief note on what they are and why they are important.

A delimiter is a “special character or string used to separate different sections of the prompt (Delimiter - Learnius, n.d.).” They help the LLM to understand which parts are instructions and which parts are data to be processed.

Delimiters include:

• Triple quotes """
• Triple backticks ```
• Triple dashes - - -
• Angle brackets < >
• XML tags <tag> </tag>
• Section titles # Role

Delimiters often go hand-in-hand with Variables. Simply put, variables hold data. The variable ‘FirstName’ holds a person’s First Name.

We use variables in our prompt. Our variables are:

• The information we want on the customer;
• The market niche.

Ottley suggests we add the variables directly after the Task.

Embedding Prompt Commands — Debunking and Discerning the LLM Output

It took me some time to write this article. Something bugged me, and I woke one morning realising what it is.

A sobering moment!

Our use case — Initial Customer Profile 'Research' — as it stands is fictitious. LLMs are brilliant at fabricating information.

I use commands to debunk and validate the LLM output. In this way I use my own discernment to decide to accept, reject or iterate further.

Even after this, I am still working with a hypothesis that is tested in a real world scenario.

I first came across the use of commands during my time at the MECLABS AI Guild  . After conducting further research, I discovered the technique I share here.

To get a more detailed output on a section of output, I use embedded 'Prompt Commands'. A prompt command is a mini process you instruct the LLM to execute.

The prompt command is defined within your prompt. There are many ways to define commands, I use the format below:

• Command name: The actual command we type in.
• Purpose: What the command is to achieve.
• Procedure: The procedure for the model to follow in executing the command.
• Parameters: The information the command requires.
• Example Usage: An example of the command in use.

There are many potential commands. I will use three.

• /validateOutput — Provide a detailed justification for the LLM's output, explaining the choices, reasons, and logic behind those choices.
• /expandDetail — Enrich the discussion with detailed explanations, background information, and personal insights.
• /citeDetail — Provide the sources of information, including detailed citations.

How to use a command:

• Our initial LLM output gives a Profile ‘The AI-driven Strategist’ with a brief paragraph on buying habits;
• We prompt /expandDetail topic="The AI-driven Strategist" depth="high".

Jump to the Completed Prompt  .

The next section is specifics. It contains the most important notes or directives to follow when executing the task.

A function of Specifics is to add new directives as we test the prompt. Do not bloat the prompt by adding ‘fluff’.

Some directives to include here are:

• Format:
  • Output your answer as a table with two headings, Attribute and Description.
  • Include a heading followed by a summary paragraph and then detailed paragraphs on each of the attributes.
• Style: In the style of Seth Goden, Ernest Hemingway, Malcolm Gladwell etc.
• Tone: Conversational tone, formal tone, professional tone etc.

Emotional Prompting (EmotionPrompt)

The research of Li et al., 2023 shows the effect on performance of adding emotional stimuli as a:

• 8% increase on simple tasks and 115% increase on complex tasks;
• 19% increase on truthfulness; and
• 12% increase in informativeness.

The EmotionPrompt is adding short phrases of emotional stimuli.

Examples of EmotionPrompts:

• For simple tasks: “This is very important to my career”; and
• For complex tasks: “This task is vital for my career, and I greatly value your thorough analysis (Liam Ottley, 2024).”

Jump to the Completed Prompt  .

Context provides the environment of the LLM. This suggests (infers) factors that influence and inform the output. Give different levels of context:

• Macro — context within society;
• Business — context relating to the business;
• System — what role does the LLM play in a system of process.

Context answers the question, “Why am I doing this?”

Context compliments Role Prompting by “further clarifying who it is, what it is doing and why it is doing it (Liam Ottley, 2024).”

Add EmotionPrompt to emphasise the importance of the LLM’s contribution — to the success of your career, business, and society.

Jump to the Completed Prompt  .

Adding examples — few-shot prompting — increases performance and fine-tunes “response tone, format, and length (Liam Ottley, 2024).”

Key points on examples:

• Examples take the form of input | output response pairs;
• Few-shot prompting serves the function of ‘in-context learning’ (learn-by-doing, learn on the job);
• It is an effective substitute for fine-tuning;
• The ‘correctness’ of the examples has no impact on the performance. Examples are not performing a ‘learning’ function, but rather informing format and structure (Liam Ottley, 2024).

Research shows adding examples increases performance by an average of 14.4% over prompts that do not have examples (zero-shot prompting).

This number is achieved with 32 examples per task. Creating 32 examples adds a time overhead not viable for most use cases. Ottley suggests using 3 - 5 examples.

In our case, we will not be adding examples. We are not looking for a ‘correct answer’, but rather a hypothesis. Our outputs will be lengthy. Adding examples, in this case, will negatively impact token use.

Jump to the Completed Prompt  .

To appreciate notes, we look at the ‘Lost in the Middle Effect’.

With long prompts, LLMs perform better (are more accurate) when important information is placed at the beginning or end (Venkatesan, 2023). Performance “significantly worsens when critical information is in the middle of a long context (Liam Ottley, 2024).

Research shows an increase of 20 - 25% in accuracy with the correct placement of important context.

We use notes to remind the LLM of essential information. Notes are in the form of a list:

• Key point reminders from the task and specifics;
• Tone adjustments;
• Prohibitions — the “do not do’s*”;
• Output format instructions.

^* A note on the “do not do’s”. How to deal with outputs, words, formats etc. we do not want. LLMs have a problem understanding negations (Rosenbaum, 2023). It has to do with the way models are trained and the ambiguity introduced with the “do not”.

Instead of “do not use the phrase ‘in the realm of’”, use “replace the phrase ‘in the realm of’ with ‘in’.”

Useful GPTs

These GPTs are useful for building your own Prompt Commands.

• Command GPT
• GPT "Prompt Command" Maker

Links and downloads:

I can help you with your:

1. AI strategy;
2. Prompt engineering;
3. Content creation;
4. Custom GPTs.

I am available for remote freelance work. Please contact me.

This article is made possible because of the following excellent resources.

The following articles are of interest.