Mental Modal Before Prompting The Right Way

A guide for prompting from my personal experience

Preparation for the prompt is 10x more important than the prompt itself

Prompting will automatically happen once you are prepared

Things to Do Before You Start Prompting

A progressive roadmap of what you want to develop,

Example you want to develop an AI powered, trip finder aggregator platform (experiences like relaxation, creativity etc)

Now this would require multiple things

The basics like your auth, databases, landing pages, UI/UX, user journeys how it would look, payment gateways, how would bookings will happen etc

Now you need to think about the most important parts which are unique and important

Data sources, how will you get the listings where does that data come from

Will you be coordinating the bookings or redirect users to the source from where the data is found

Then even if you want to have AI powered search, you can't build the AI powered search and suggestions until the base platform is in place, I would leave this aside for now and develop the base platform

Then I will implement the AI search and suggestions into the platform, after deciding the UX

Thinking about your prompts and how to access the knowledge the AI models have

Now developing an aggregator platform is something the AI can do rather easily

Why?

Because there are a lot of aggregator platforms in the world already -> which implies there is a lot of opensource/code on which these models have already trained

The AI search and suggestions, that feels new

Not a lot of products with AI search and suggestions

Not a lot of code on which the model has been trained already to give you the perfect output

You want to use the latest standards and best practices, which in most cases are available on dates which are beyond any LLM models training cutoff date

The search algos might be easier, but AI search, when to embed data, how to embed data, how to chunk it, how to search it, RAG or Combined, or simple search and feed results to AI and generate suggestions?

Now for this what I do is create docs or even ask Claude Code to research available docs, and save important patterns, methods to use the LLM APIs, how to handle responses, how to use tools etc in docs within folders

Which helps me guide the model easily to read and implement the methods correctly

So in a nutshell, it's fairly logical and simple, but you need to know what you are doing, play around with models yourself to figure out what limitations there are, what code it might already be trained on a lot, and whatever is new, it's safer to create your own docs, or ask the model to research and create docs for more accurate outputs and save a lot of time.

Another quick example to clarify more little nuances

In early days like 2 years ago, when you asked models to create websites, most of the time the button placements, like the most basic thing in a navigation bar would be wrong

Then I would tell the model to place it in the right corner of the page

The model will place it in the left, and then I will tell you placed it in the left and then it would correct it

This made me realise I am talking to an API coming from a server sitting somewhere, it has no sense of direction, size or color, it can tell left right, north south up down it's all relative

Then I used to prompt with references: place the button to the right side of the logo which is on the left end of the nav bar

The output? - Perfect in the first time

Now with latest models you don't see this basic issue anymore, but there are always new gaps when you solve a problem, you find new problems, the models are not smart or intelligent or not actually "think" - it's more of logical ways we are forcing them to access the vast amount of knowledge they have, and that's what is great once you assume the model is dumb, but the most knowledgeable entity in the world, you need to figure out how to guide it and access the knowledge it has -> For that you need to become smarter, which you can but you can't become more knowledgeable than a model

Other Good Practices

Instead of technical details,

Your prompts should focus more like a PRD from a product manager

Add user journeys, for the feature you are developing

even if simple journeys, but do that your output will be way better.

The expected system behavior, the tradeoffs of you are doing to implement this

The not to do, overengineering stuff, why we are implementing only this and what is excluded so the model never generates unnecessary extra functionalities, which might be the very logical extensions of your feature but you might not need it, probability will force the model to implement it even if you don't want to

Example

You want to setup login and signup

Now many times for a quick MVP you won't even need password reset

You only want login with Google, no email OTP or magic links

If you don't specify this, your login/signup will like the most available in the industry which have social + email/username password logins with password reset pages etc