Using AI to Plan Weight Loss: A Practical Guide

Two ways to approach weight loss with AI. Approach one: type “give me quick weight loss tips” into ChatGPT and get the same generic list you’d find on any health blog – drink water, eat protein, sleep more. Approach two: use the model as a planning partner that adapts to your actual schedule, food preferences, and constraints. The second one works. The first one wastes the tool.

This guide covers the second approach – how to prompt ChatGPT or Claude to build a realistic, personalized weight loss plan, and where these models quietly get things wrong.

Why generic tips fail – and what AI actually adds

Generic tips don’t fail because they’re wrong. They fail because they’re not yours. “Eat more protein” is true. It’s also useless if you don’t know how much, from what, at what time, or how to fit it into a Tuesday when you skip breakfast and eat lunch at 2pm from a corner shop.

Think of an LLM like a very fast, tireless meal-planning assistant who has read every fitness book ever written but has never met you. The gap isn’t knowledge – it’s context. Your job is to give it that context before asking for anything. Do that, and the output stops being generic advice and starts being something you can actually use on a Wednesday night when you’re tired and the fridge is mostly condiments.

Three prompt patterns that work

Skip one-liner prompts. These three patterns produce plans people actually follow. The third one came from two weeks of the first one failing – stumbled onto it by accident.

Pattern 1: Constraints first, plan second

Tell the model everything about your life before asking for a plan. Opposite of how most people prompt.

I want to lose about 6 kg over 3 months. Here's my situation:
- I'm 34, sedentary desk job, 78kg, 175cm
- I cook dinner 4 nights/week, eat out 3 nights
- I don't eat breakfast, coffee only until 11am
- I hate meal prep on Sundays
- Budget: roughly £60/week for groceries
- I lift weights 2x/week, no cardio

Build me a weekly eating pattern (not a rigid meal plan) that fits this. Assume a 400-500 kcal deficit. Give me 5 dinner options I can rotate.

Constraints handed over, not tips requested. You’ll get a rotation you can actually use – not a Pinterest meal plan with 14 ingredients you’ll never buy.

Pattern 2: The “argue with me” follow-up

After the first plan lands, paste this:

Now play devil's advocate. What are the three most likely reasons this plan fails for someone like me? Then revise it to address those failure modes.

LLMs are trained to be agreeable. Ask them to critique their own output and you get a much more honest second draft. This is the single most useful prompt in the whole workflow – the first plan is a sketch, this turns it into something stress-tested.

Pattern 3: The weekly check-in loop

Every Sunday, paste last week’s rough food log and ask the model to spot patterns. Turns out Claude’s 200K token context window (on paid tiers, per Anthropic’s documentation) makes a real difference here – paste four weeks of logs at once and get pattern analysis across the whole stretch. ChatGPT handles this too, though context limits vary by tier, so long multi-week pastes may get truncated.

Pro tip: Don’t ask “how am I doing?” – you’ll get vague encouragement. Ask “what’s the one thing I’m consistently getting wrong?” Specific question, specific answer.

Where the AI quietly gets it wrong

Hit all three of these before the workflow clicked.

Hallucinated calorie counts. Ask ChatGPT how many calories are in a Pret chicken sandwich and it gives you a confident number. Often invented. Research on LLM hallucination (documented in academic literature) shows models fabricate specific numeric values – branded food items are especially unreliable since exact nutrition data may not have been in training data. Cross-check anything specific against the actual product page.

Assumed US portion sizes. Both Claude and ChatGPT default to American serving sizes and imperial units unless told otherwise. If you’re in the UK, EU, or elsewhere, the plan quietly overestimates portions. Fix: state your country and preferred units in the first prompt.

“Diet coach” custom GPTs. Most in the GPT Store are the base model plus a system prompt – OpenAI’s own documentation confirms custom GPTs run on the same underlying model. The nutrition accuracy isn’t improved. Just the branding.

What 8 weeks actually looks like

A 500 kcal/day deficit produces roughly 0.45 kg of fat loss per week – that’s the standard nutrition science figure, and it’s where most AI-generated plans anchor (basic nutrition science, consistent with NIH guidance). Over 8 weeks: around 3.5 kg, assuming adherence. Adherence is where most plans die.

The AI workflow doesn’t change the biology. What it changes is the friction. Wednesday goes sideways – you paste the situation into the chat, get a revised Thursday in 30 seconds, and keep going. That’s the actual value. Not the plan. The recovery from breaking the plan.

The catch: when not to use this

Hard stops:

• BMI over 40 or under 18.5 – both ChatGPT and Claude redirect to a doctor at these ranges, correctly, per OpenAI’s usage policies on medical advice. Good policy, rarely mentioned in AI-diet tutorials.
• Diagnosed eating disorder history. LLMs aren’t clinically supervised and can reinforce restrictive patterns without flagging them.
• Medications that interact with diet (diabetes, thyroid, blood thinners). The model doesn’t know your prescriptions and won’t ask.
• You want accountability. AI never notices you ghosted it for a week.

FAQ

Is ChatGPT or Claude better for this?

Claude. The longer context window handles multi-week logs without truncating. ChatGPT is faster for one-off prompts, but for this workflow specifically, Claude wins.

Do I need a paid plan?

For the first few weeks, no – the free tier of either model handles the basics. The friction starts during Pattern 3 when you paste a long food log and hit the message cap mid-analysis. At that point, upgrading makes sense. Pricing changes frequently; check the current pages for ChatGPT and Claude directly before subscribing rather than trusting any number you read here.

Can the AI accurately count my calories?

No. And this is worth sitting with for a second, because it’s the most common trap people fall into. The model will give you exact calorie figures with zero hesitation – and those numbers are often fabricated, especially for branded or restaurant items (see the hallucination research linked above). Use MyFitnessPal or Cronometer for actual logging. Use the LLM for pattern analysis on top of that data. Two different tools, two different jobs.

Start here

Open ChatGPT or Claude now and paste Pattern 1 with your real constraints filled in. The plan you get in 90 seconds is the starting point everything else builds on.