If you’ve ever watched The Try Guys, you know how entertaining it can be to try and recreate complex recipes without a written list of ingredients and steps to follow.
And if you’ve ever tried to ask your favourite AI Chat tool to work on a complex task for you, you know how not entertaining it can be to try and correct its mistakes.
It reminds me of a mistake I made early in my career, when I got my first junior marketing role. I sent an invitation email out about an event to our subscribers, forgetting to include an important detail — sponsorship logos. But, I had:
- No context such as previous examples of emails they had used, since I was also implementing an entirely new email tool,
- No instructions beyond “make an announcement email”, and
- No one to verify I had done it correctly since I was the only marketing/communications team member
In a way, I myself was a bit like a Chat LLM who had been asked to produce something professional. I was competent enough (I had a marketing degree), but I wasn’t really trained or instructed properly. So, the quality of my work was limited by my own ability to produce — I needed mentorship and guidance to get better, and eventually, I got those support systems.
AI works in a similar fashion, except that it needs continuous mentorship and guidance. As of right now, it cannot simply walk off into the world after gaining enough experience and start its own company (like I eventually did). Right now, you can think of it as your “smartest intern” who doesn’t really know what’s going on but will work its butt off to prove itself to you. Useful, but not a standalone resource.
There are many elements that can go into AI “cooking” and I’d like to cover some of them to help you think more about what you’re trying to accomplish. However, I’m going to keep the technical jargon to a minimum, keeping with the “food” format, because if you’re a small business owner, I don’t want to add more to your daily headaches. I want to provide clear and actionable strategies that you might incorporate into your work.
Prompts (the recipe)
A prompt is an instruction — anything as simple as “tell me about today’s weather”. A prompt can be used in multiple “ways” — think like, roasting, frying, or boiling. Depending on what you’re trying to accomplish, you’ll want to prompt in different ways.
If your interaction with the AI is just speaking to it (via Chat), it’s generally ok to just write out your constraints in plain text: “You are acting as a consultant, I do not want you to be too agreeable, I want you to challenge my assumptions, and you need to cite your sources with data from the past 12 months”. A line like this will set expectations.
But, if you want an AI to help translate data in your CRM from one place to another, you’re going to have to work with a more technical tool — something that can safely send the data, like through an API. An API is sort of like a very secure email address that you need a password to send email to. And instead of using an email tool like Gmail or Outlook, you use code.
Then, you might build code that will give certain instructions in the prompt. To understand these instructions and what they do, I’ve tried to make some comparisons:
| Instruction | Purpose | Role in “recipe” |
|---|---|---|
| A system prompt | Expected behaviour, necessary constraints, and the output format | The steps |
| User + assistant pairs | Referred to as “few shot” examples — this loosely or more closely tells the AI what your desired output looks like | A picture of the dish |
| Temperature or top_p | Controls creativity vs consistency (from 0.0–1.0) | How creative should the result be? Can “ingredients” be added based on vibes, or does everything need to be measured precisely so it always comes out the same way? |
| Max_tokens | Limits response length | How much are we making? And therefore how much will the ingredients cost? |
Find some great resources for learning more about prompt engineering here: promptingguide.ai
Many AI resources focus heavily on prompts because it is undoubtedly THE single largest point of failure for using AI effectively! A good prompt, coupled with good data and a well-orchestrated system can take care of incredibly complex tasks. Good data and the rest of the system alone are not enough. Conversely, the same can be said about prompts.
Therefore, there are still other pieces you need to know about.
Prompt Chaining (recipes in sequence)
We won’t get too in depth about this, but this is like doing a bunch of different recipes in sequence to result in a final meal. Think your most elaborate holiday meal — for Christmas, Thanksgiving, Eid, Passover Seder, the Lunar New Year, Diwali… you prepare some of the side dishes the day before, the centrepiece gets started roasting in the morning, you need to make sauces and breads to dip or hold food, and you need a show-stopping dessert. You can also use many prompts, one after another, to create a final product of work that is truly worth admiration.
Function calling / Tools
LLMs, when given access to certain tools or functions, can decide to use them based on your prompt.
If you’re a bread baker, you know that using a bread machine or stand mixer produces more reliable results than kneading by hand (unless you’re Paul Hollywood).
Tools allow LLMs to access data elsewhere — such as up to date weather information, or your customer data — so that they’re less likely to hallucinate, and more likely to produce something reliable. Tools can be built-in to your LLM (such as a way to search Google for missing information) or can be something you set up to work with your own technology stack as part of your “cooking”.
RAG
This stands for Retrieval Augmented Generation. It means you have a library of content that the LLM can get more information from.
One of my favourite cookbooks is “Mastering the Art of French Cooking” written by Simone Beck and Louisette Bertholle, both from France, and Julia Child, an American. Throughout its many recipes, you will find references to other pages — sauces, methods, ingredient definitions (such as the “bouquet garni” which many outside France have never heard of). With only one page of this cookbook, you may not be able to realize your vision. With the use of the entire cookbook, you should master French cooking in no-time.
However, with an LLM it goes further — RAG also allows you to customise your AI of choice without expensive (and environmentally impactful) retraining. Along with the knowledge the AI already has, it can also access databases of knowledge from your organization that are “vectorized” (a process that increases efficiency in how the information can be used by the AI) and it can be easily updated as you gain more knowledge or create more content.
Workflows
My best way to put this simply is that it is like a pizza vending machine. Like the one outside the CHL. If you didn’t know about it — now you know. You’re welcome.
It’s more complicated than that, but basically: a workflow can execute a set of instructions automatically to create either a single output or a variety of them. However, you more or less understand what you want the final output to look like.
- First, there is a trigger (someone orders the pizza).
- Then, a pre-determined set of steps follows (the pizza is assembled and cooked).
- Then, the task is complete (the pizza leaves the machine).
I can help you build a workflow to take care of tasks that are slowing you down. I also have pre-built workflows you can demo and keep for free which you can access in the “Workflow Assistants” tab in my navbar. Or, just get in touch.
MCP
Standing for “Model Context Protocol”, this is one of the more complicated pieces of AI-kitchen-wizardry.
Created by Anthropic, it has been adopted by much of the AI industry as a standard for how to connect different tools and systems to LLMs.
It’s sort of like the “mise en place” of AI. Every ingredient and tool in its place, prepared the correct way, so you don’t have to hunt for it — you can just do what you need to do without stopping to fix/prepare another element.
Creating an MCP server of your own is possible, and a lot of businesses want to have this ability because they want to take on many projects with AI. The reusability of an MCP server, therefore, is quite useful. However, if you’re not a developer, there are ways to use MCP without advanced technical skills. Your CRM provider, for example, might have an MCP server you can use as their customer. There are also 3rd party MCP server integrations for some CRMs who don’t yet have their own. These I would be more cautious of — but if they’ve been approved by your CRM provider, you may proceed diligently.
You still may need a helping hand, whatever you choose. This is fine. You also probably don’t need an MCP server if you’re just starting out. Start with tool integrations, low-code workflows (n8n or similar tools), and just getting beyond Chat before you tackle MCP.
Mixture of Experts (MoE)
Ever heard of a brigade style kitchen? This is where multiple chefs have defined roles — the executive, the one who makes the pastry, someone for the sauces, someone for deep frying — you name it.
This is done for efficiency. Similarly, a MoE AI model employs multiple “experts”, as in, chunks of models, to solve problems. Each “expert” has been deemed to be of a certain aptitude, and when that aptitude is NOT called for, they do not “light up”.
This increases efficiency and is actually a way some are trying to decrease the impact of AI’s energy use on our world. If you ask a question about mathematics, the Julia Child Trivia section of the AI’s “brain” doesn’t need to light up. Unless of course, you’re trying to convert her recipes to different measurement systems! But this saves significant energy and can actually result in better outputs, too.
You can read more about this from Datacamp.
AI Agents
Agents are autonomous and can make decisions. This makes them useful, but dangerous. Think a cook who is out of tahini, and throws peanut butter in a dish instead — not knowing the person who is eating it has a deadly peanut allergy!
There is no complete definition of an Agent that everyone can agree on. However, the main qualities seem to be that they operate independently, and they can make decisions. This makes them different than a workflow with steps that include connection to an AI model API. They do not exactly need to follow the workflow’s steps and may go out of “bounds” to accomplish tasks if not effectively constrained.
The cook needs to know that the person they are cooking for is allergic to peanuts. They need to know that the consequences are dire. Your AI model needs to know it cannot overstep certain bounds, such as giving false information to a customer, deleting data, or sending emails it is not approved to send.
To take a more serious tone than the rest of this blog post: Allowing an AI agent to have too much agency over your business is a recipe for disaster. Keeping an eye on your Agents and their decisions is important for your credibility.
Finally: AI Harnesses
This is beyond what you can accomplish in a home kitchen — the chef is the AI model, who is given the meal orders and access to tools, knowledge, context, quality ingredients, and methods that home cooks don’t have. They must ensure ingredients are ordered as needed, that their staff are following their orders, that their tools are in good condition; that everything is orchestrated the way you’d expect a Michelin star experience to work.
With an AI Harness, you may have multiple tools, persistent memory, and verification steps that make a process that truly seems like magic.
It’s a level of maturity most organisations do not yet have, so don’t let it intimidate you — but know that one day, you could be working at this level, even if you have a very small team.
I hope this post was useful to someone out there who wants to expand their knowledge of AI but isn’t in a technical role.
I want to make AI accessible to everyone, especially small business owners in Luxembourg who may not have the time to get an AI certification, or figure out how to use AI effectively to help increase their team’s efficiency and growth. This includes custom workflow creation, tool implementation, support, training, and more.
If you want to talk more about how AI can help your business grow, please get in touch.