3 years ago, Open AI introduced us to a baby called ChatGPT.
This baby wasn’t human. But the abilities it possessed were beyond what any single human could ever possess. It could write stories, poems, songs, introductions, and profiles. It could help with your homework essay, your thesis, that report due on Monday, or even the software running your home electronics.
Not only was this child a genius with infinite knowledge and patience, it was also:
-
Omnipresent (in your nearest laptop/mobile)
-
**Omniscient (Well, everything that the internet does) 3. Multilingual (**it understood over 100 languages, to a certain extent)
And yet, most of us don’t fully understand how capable it is,
I know I didn’t.
I used to type in a random sentence and expect magic to happen, which sometimes it did, but it also didn’t happen a LOT of times.
No — that’s not what I’m saying..
Most people already know how to use AI. The difference is knowing how to get consistently great results from it. All I want is for you to know how to get the best results, whenever you want. And all we need to is structure our prompts just a tad differently.
The convenient thing is, this works for all subsequently introduced babies as well. So, let’s learn how to do that by learning different types of prompting.
This is what most of us have been doing all along. You simply type your request and let the AI figure out the rest.
The model decides how to interpret your request, what reasoning process to use, what assumptions to make, and how to structure its response. Sometimes this works brilliantly. Other times, the AI fills in gaps you never intended it to fill.
Prompt:“It’s my best friend’s wedding. Help me find a gift for him and his fiancée.”
The AI makes all the decisions here — tone, length, wording, and even assumptions like it could be Jeff Bezos asking this, money is no bar.
When you prompt the AI for a query and in your prompt, you share an example that explains the relationship between the query and it’s response in a direct or subtle manner.
It teaches the model the tone, format and classification logic without explaining it.
“Help me find a wedding gift based on what a couple loves.Prompt:Example:They love cooking -> A premium Damascus steel chef’s knife set.It’s my best friend’s wedding. Help me find a gift for him and his fiancée. They love to hike a lot, so suggest something related that helps them enjoy it more.”Now solve this:
No, I’m not being naughty here. This is a legitimate strategy.
In this type of a prompt, you assign the AI a specific identity, profession, or viewpoint. This changes the model’s vocabulary, assumptions, and formatting.
“It’s Your Best friend’s wedding, find a gift for him and his fiancée who actively trek a lot. You have been to a couple treks with them. A couple times, the weather restricted them from fully enjoying it due to either dehydration or lack of food, or something else. Suggest something that makes it easier for them to enjoy trekking more”Prompt:
It’s setting up context and user’s viewpoint. It changes how AI approaches the response.
Strictly defining the output structure. The response produced via such prompts is called a structured output. This is vital when building apps on top of LLMs. The data format needs to be consistent. If not, the response may breaks the application’s processing pipeline.
“It’s my Best friend’s wedding, help me find a gift for him and his fiancée. Share the items in a JSON format ONLY. Make sure to include the brand name, prices, merchant site availability, upcoming discounts and most importantly, a rating that describes how useful it is during a trek session”Prompt:
The LLM knows what we want, and it gets to work!
There’s also a slight disadvantage that this type of prompt introduces. We want the AI to follow logic and return the response in a specific format. But we can’t control which task gets more weight. AI may decide to use more computation to get the format exactly right, but in the process, it may compromise on it’s logical reasoning.
A massive breakthrough discovered by researchers. Simply appending the magic phrase ** “Let’s think step by step”** to a zero-shot prompt forces the model to generate a chain of thought on the fly, drastically improving logic without needing hand-crafted examples.
The goal here is to force the model to slow down and map out its logic before giving you the final gift list, reducing the chance of generic answers.
Prompt:“It’s my best friend’s wedding, help me find a gift for him and his fiancée who are avid trekkers.Let’s think step-by-step:First, identify the biggest comfort and safety challenges couples face on long treks. Second, brainstorm premium gear solutions that address those specific challenges. Finally, recommend the top 3 best wedding gift options based on that breakdown.”
To fix the issue we face in the previous type, we combine Type 4 (Constraint) with Type 5 (Chain of Thought). Instead of telling the AI to hide its thinking, we give it a specific place inside the JSON to do its thinking before it gives the final answer.
{ "thought_process": "The couple needs premium gear for intense mountain treks where they frequently face unexpected heavy rainstorms. A standard gift won't cut it. They need something ultra-lightweight, highly durable, and completely weatherproof to keep them dry and comfortable...", "brand_name": "MSR", "gift_item": "Hubba Hubba NX 2-Person Backpacking Tent", "price": "$549.95", "merchant_site_availability": ["REI", "Amazon", "MSR Gear"], "upcoming_discounts": "15% off during the upcoming REI Anniversary Sale", "trek_usefulness_rating": "9.8/10 (Absolute lifesaver for heavy rainstorms and ultra-lightweight packing)"}
But you don’t need to get into that right now
In this type, instead of just giving examples of gifts, you give an example of the reasoning process you want the AI to follow.
You provide a few-shot example that includes the step-by-step reasoning leading to the answer, not just the answer itself. By teaching the model to output its intermediate steps, its accuracy on complex tasks skyrockets.
Prompt:“It’s my best friend’s wedding, help me find a gift for them. They love trekking. Follow the reasoning logic shown in the example below to give your recommendation:Love beach camping, hate sleeping on uneven sand, love cooking.Example:
- Couple Profile:
-Reasoning:Regular tents are too hot. They need elevated comfort and utility. A premium double-sized camping cot solves the sand issue, and a portable marine-grade cooler elevates their beach cooking experience.
-Gift Recommendation:A Yeti Tundra Cooler or a Kamprite Double Tent Cot.
Now, solve for this couple:They love intense mountain trekking, frequently get caught in unexpected heavy rainstorms, and prioritize ultra-lightweight packing.”
- Couple Profile:
A highly advanced technique where you prompt the model to generate multiple possible reasoning paths (branches), evaluate the promise of each path, and then decide which one to follow to the conclusion. It mimics human brainstorming.
Prompt:“It’s my best friend’s wedding, help me find a trekking gift for him and his wife. To find the perfect item, imagine three independent outdoor retail experts brainstorming.
-Expert 1explores the ‘Ultra-lightweight Tech’ angle.
-Expert 2explores the ‘Camp Comfort & Luxury’ angle.
-Expert 3explores the ‘Extreme Weather Survival’ angle.
Have each expert propose one premium gift idea and critique the other experts’ ideas. Finally, have them reach a consensus on the single best gift.”
This forces the LLM to act like a committee of experts, exploring multiple paths, judging them, and picking the winner.
Prompt chaining aims at breaking a massive task into a sequence of smaller, specialized prompts, where the output of Prompt 1 becomes the input for Prompt 2. This one is best explained by an example:
Prompt 1:“My best friend and his fiancé love long-distance trekking. Generate a list of the top 5 most common gear failures or pain points that happen to couples on 3-day mountain hikes.”
(You copy the AI’s list of pain points, then run Prompt 2)
“Based on these specific trekking pain points:Prompt 2:[Paste Output from Prompt 1 Here], act as a luxury wedding concierge and recommend one high-end, durable wedding gift that solves each problem permanently.”
Easy, right?
You give the model explicit hints (or guardrails) to steer the creative direction without giving it a rigid template. Just like guided meditation, this is guided generation. Wherever you go, Hutch dog follows!
“It’s my best friend’s wedding, help me find a gift for him and his to-be wife. They are hardcore mountain hikers.Prompt:
Focus the suggestions exclusively on items with a lifetime warranty, products made from sustainable/recycled materials, and gear that is specifically designed for use byDirectional Hints:twopeople simultaneously.”
(I can say for sure that this name was not coined by Meta, wink wink)
A technique primarily used for AI agents. You prompt the model to alternate between “thinking” about what to do next, and “acting” by outputting a command to use a tool.
It’s like providing a blueprint for AI Agents. You prompt the model to alternate between “thinking” and “acting” (using tools like a web search or calculator) to achieve desired response.
“You are an AI Shopping Assistant with access to a Google Search tool. Help me find a wedding gift for my best friend and his wife who love trekking. You must alternate between Thought (why you need information), Action (searching the web), and Observation (what you found) until you find a unique, highly-rated item.Prompt:
User:‘They love hiking but always complain about getting dehydrated or freezing at night. Find a premium gift available online right now.’”
So, there you have it. You are officially out of excuses for getting bad responses from an LLM.
You don’t need to use a highly advanced Tree of Thoughts prompt just to write a basic email, but knowing how to shift gears when the AI gets stuck is your new superpower. Treat your prompts like a blueprint, not a slot machine.
Happy prompting! (And if you’re building an app with ReAct, just remember: Meta had nothing to do with it).
Sources:
Production App Guides (JSON/Chaining):
No, that is not how you prompt! was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.