Nutrition Tracker Part 1: Understanding LLM Generic Answers

🔥

Series: Learn ChatGPT by Building a Nutrition and Activity Tracker

In the introduction to this series, I shared how I started using ChatGPT to build a personalized nutrition and activity tracking system.

Today, it tracks food, activity, workouts, steps, estimated calorie burn, bodyweight, and trends. It also helps generate dashboards, calculate custom metrics, analyze patterns, and provide feedback based on my goals.

But it did not start there.

It started with a simple food log.

The Experiment:

I opened a new ChatGPT conversation and entered something like, today I ate, and entered my food.

Breakfast:

2 scrambled eggs
2 slices wheat toast

Lunch:

Turkey sandwich
Small apple

Dinner:

Grilled chicken breast
1 cup white rice
Mixed vegetables

Snacks:

Protein bar
Handful of potato chips

Activity:

7,500 steps
30-minute walk

Try it. Simply copy and paste the list above into a standard chat, and follow with the prompt, "How did I do today?"

In my case, the response was reasonable. ChatGPT told me I had a fairly balanced day, included some protein, got in some movement, and made some generally healthy choices.

That was fine. But it was also generic.

So I asked a follow-up question. Was I in a calorie deficit? And ChatGPT essentially said, I do not have enough information to know.

The model did not fail.

It simply did not have enough information.

At that point, ChatGPT had no context, and therefore did not know:

My weight
My height
My goals
My calorie target
My protein target
My BMR

So the response was very general.

It could comment on the food.

It could estimate calories.

It could provide broad health advice.

But it could not tell me whether the day was successful against my actual goals.

A generic prompt produces a generic answer.

A personalized system produces a contextual answer.

At this stage, I was NOT using:

Projects
Project Instructions
Memory

I was simply chatting with ChatGPT.

And the results reflected that.

This is one of the most important things to understand about ChatGPT and large language models.

They can only reason from the information available to them. Without context, the model will usually give you an answer that is broadly correct, generally useful, and safe for the average person.

But most of us are not asking average questions. We want answers that apply to our goals, our constraints, our habits, our data, and our definition of success.

That requires context.

What Happens Next:

The next step simple. We do not need a complicated prompt. We need to give ChatGPT a better operating environment and personalized information.

So, we need to create a dedicated Project and add a few basic instructions, and a few pieces of information: