ChatGPT Weight Loss Prompts That Do Something Useful

A prompt changes the shape of the answer, not where the numbers come from. These ChatGPT weight loss prompts are written to produce output you can check: gram weights, stated assumptions, and arithmetic you can add up yourself. Copy them, and read what each one is bad at.

The one principle

Most weight loss prompt collections are variations on "act as an expert nutritionist and create a personalised plan". They produce confident, fluent output. They do not produce output you can audit, because nothing in the reply tells you where a number came from or what was assumed.

The useful prompts do two things instead:

• Ask for structure, not facts. A model is reliable at arranging, comparing and explaining. It is unreliable at recalling what a specific branded food contains.
• Force the assumptions into the open. Any weight it guessed, any conversion it made, any figure it is unsure about, should appear in the answer rather than under it.

Every prompt below is built that way.

Setting a target

I am 34, male, 82 kg, 180 cm, lifting four times a week and otherwise
at a desk. I want to lose fat.

Give me a maintenance estimate and a moderate deficit target. Show the
formula you used and the activity multiplier. Then tell me what would
make this estimate wrong for me specifically.

Good at: the arithmetic. Resting expenditure formulas are well represented and it will apply one correctly.

Bad at: knowing whether the formula fits you. It is a population estimate, and two people with identical inputs differ by several hundred calories a day. The last sentence of the prompt is what makes it useful, because it surfaces the caveats instead of hiding them.

What to do with it: hold the number for two weeks, track a weekly average weight rather than daily, and adjust from what the scale did.

Planning a day

Build me a day at 2,150 kcal and at least 175 g protein.

Rules: every food gets a gram weight, no "servings". Say whether each
weight is raw or cooked. Give per-meal subtotals and a daily total, and
make the total the sum of the subtotals. End with a list of any figure
you are not confident about.

Good at: hitting the target arrangement and respecting constraints.

Bad at: the per-food figures, especially branded items. The final instruction surfaces exactly those.

Logging a meal

Log this: 220 g plain skyr, 60 g dry oats, one medium banana,
15 g peanut butter.

Tell me the gram weight you assumed for the banana. For each item, say
whether your figure is for the raw or cooked form. Then total it and
show the 4/4/9 check: protein and carbs at 4 kcal per gram, fat at 9,
compared with your calorie figure.

Good at: the totalling, once the inputs are right.

Bad at: "one medium banana". Something becomes a gram weight invisibly unless you make it say so.

The 4/4/9 line is the highest-value instruction in this whole post. It catches figures where the macros and the calories were produced independently rather than derived from each other.

The weekly review

Here is my log for the last 14 days. [paste]

Give me, in this order: how many days out of 14 I actually logged, the
average calories for each week, the average protein, and the weekly
average weight for each week. Then say whether to hold or adjust, and
explain which of those four numbers drove your answer.

Good at: reading a text log and summarising it. This is text processing, which is what the model is actually for.

Bad at: nothing much, which is why this is the prompt with the best value-to-risk ratio here.

Why the order matters: days logged comes first deliberately. If you logged four days out of fourteen, the averages cover the days you chose to record, which are systematically the good ones, and nothing else in the review means anything.

Diagnosing a stall

My weight has not moved in four weeks. My logged average is 2,150 kcal
and my target is 2,150.

Do not tell me my metabolism is damaged. Give me the two most likely
explanations, and for each one, a specific test I can run in the next
four days that would distinguish it from the other.

Good at: producing the right two candidates, which are almost always under-reported intake and reduced expenditure, plus a sensible test.

Bad at: it will sometimes reach for metabolic adaptation first because that is the more interesting answer. The first line of the prompt blocks that.

The answer you want: weigh everything strictly for four days. If the logged total rises, it was under-reporting. If four strict days land on the number you had been logging, it was expenditure. The full version is in the fat-loss tracking post.

Eating out

Dinner out tonight at an independent place, chicken thigh curry with
rice and a naan. Give me a low and high estimate rather than a single
number, assume restaurant-generous cooking fat, and then rebuild the
rest of my day around the midpoint while keeping protein at 175 g.

Good at: the planning half, which is the half that matters.

Bad at: the estimate, unavoidably. Restaurant portions vary more than home portions and nobody can see the oil. A range is a more truthful log entry than a confident single figure. More on this in estimating restaurant meal macros.

Fixing a protein shortfall

My last seven days averaged 148 g protein against a 175 g target.
Here are the foods I actually eat: [list].

Rank them by protein per calorie, not per 100 g. Then tell me the
smallest change to my current days that closes the gap, and what it
costs me in calories.

Good at: the ranking and the trade-off arithmetic. Protein per calorie is the right metric in a deficit and most people ask for the wrong one.

Bad at: the underlying per-100 g figures, as always. Since this prompt operates on a short list of foods you eat regularly, it is worth verifying that list once against labels and reusing it.

The shopping list

Turn that week into a shopping list. Group by aisle, give total grams
per item, then convert to the pack sizes a shop actually sells and show
me the difference between what the plan needs and what I have to buy.

Good at: the summing and grouping.

Bad at: nothing much, and the last clause is the valuable part. The plan needs 820 g of yoghurt and the shop sells 500 g tubs, so you buy 1,000 g. That 180 g either goes to waste or gets eaten, and if it gets eaten your week quietly ran over. Most tools hide that gap. The full version is in building a grocery list from a calorie target.

Keeping these somewhere

Re-pasting prompts is the reason people abandon this after a fortnight.

Put your profile, your targets, your food list and your house rules in one document, and start from it rather than retyping context every session. In ChatGPT that is a project instruction or a saved custom instruction. In Claude Code or Cursor it is a markdown file in the working directory that the client reads automatically.

The rules worth keeping permanently are the ones that constrain behaviour rather than request content:

- Never quote a macro figure without saying whether you looked it up
  or produced it from general knowledge.
- State the gram weight you assumed for anything I describe in words.
- Make daily totals the sum of the meal subtotals.
- Append to the log. Never rewrite earlier entries.

Those four do more for the quality of the output than any amount of prompt polish, because they apply to every reply rather than one.

Prompts that look good and are not

"Act as a world-class nutritionist with 20 years of experience." Role-play changes the tone, not the source of the numbers. It also makes the output sound more authoritative, which is the opposite of what you want.

"Give me a 30-day weight loss challenge." Produces a motivating wall of text with no mechanism for adjusting when reality diverges in week two, which it will.

"What should I eat to lose weight fast?" Invites exactly the advice that makes adherence worse, and an aggressive deficit is the one place where an agreeable assistant does real harm.

"Analyse this photo of my meal." Different problem with different failure modes, mostly portion estimation. Covered in food photo to calories.

"Track my calories for me from now on." It cannot. There is no memory of your log between sessions unless you supply one, so this produces an agreeable yes followed by an empty history. Keep the log in a file and paste it, or use a client that reads one.

The prompt to run on everything

Which of the numbers in that answer came from a database, and which
did you produce from general knowledge?

The honest answer is all of them came from general knowledge, because no lookup happened. Running this occasionally keeps you calibrated about what you are actually holding.

Before you use any of these

If you have a diagnosed condition, a therapeutic diet, a history of disordered eating, or you are pregnant or breastfeeding, weight loss prompts are not the right tool and a qualified human should be involved. Rapid loss targets and aggressive deficits are where AI-generated plans do the most harm, because nothing in the conversation pushes back.

Getting real numbers underneath the prompts

Every prompt above is damage control around one root issue: the model is recalling food composition rather than consulting it. The structural fix is to attach a tool that queries a catalog, so the assistant searches, gets a row, and scales it to your gram weight.

That is this nutrition MCP server. It authenticates with an API key header and we have verified it with Claude Code and Cursor, not inside ChatGPT, since browser sign-in is not enabled. Tools and arguments are in the docs and the plan is on MCP pricing. Building software instead? A REST plan.

What we have not measured

No test of how much these prompts improve accuracy against weighed food. They are designed to make output checkable, which is a different and more modest claim than making it correct. No accuracy figure appears in this post because we have not produced one, and we are not going to borrow one from another company's blog and present it as though it applied here.

Related

• Can ChatGPT count calories?
• ChatGPT diet plan: what it does well
• Running a cut: the weekly loop and what drifts
• Calorie tracking in Claude, a full day logged