--- title: "AI Nutrition Coach for Claude and ChatGPT: the Nutrition Plan Skill" description: Free nutrition plan skill for Claude and ChatGPT with safety questions, lower limits and sources. How to set it up and what the test showed. url: https://www.thefitfuturist.com/en/training-analysis/nutrition-plan-skill/ locale: en author: Christopher Klenk published: 2026-10-29T18:45:26.305Z modified: 2026-10-29T18:45:26.305Z --- # AI Nutrition Coach for Claude and ChatGPT: the Nutrition Plan Skill Do you want Claude or ChatGPT to build a nutrition plan that fits you, without crafting the perfect prompt every time? That is why I built the nutrition plan skill. It gives the model fixed rules: first the questions about your health and eating habits, then the numbers, each with a source. The reason was the mistakes I kept running into with ChatGPT and other AI models, especially around supplements. The skill is free for Claude, and there is a separate version for ChatGPT. If you would rather write your own prompts, my [guide to creating an AI nutrition plan](/en/training-analysis/create-ai-nutrition-plan/) shows you how. :::tldr The nutrition plan skill asks safety questions about your health and eating habits first and only then does the math. Every number carries a source tag, supplements are ranked by the strength of the evidence and come after food. Claude gets a ZIP file, ChatGPT gets project instructions plus one reference file. The skill was tested with 36 test personas on Claude, 10 on a GPT model and four real runs in the ChatGPT app. You should still check the sample days yourself. ::: :::node[jumpButton] { "label": "Go to the skills for Claude & ChatGPT", "target": "download-and-set-up-the-nutrition-plan-skill", "description": "Free: a ZIP file for Claude, a separate version for ChatGPT." } ::: ## What the skill does differently from a prompt A prompt describes what you want. A skill also brings fixed rules that the model applies to every plan: in which order it asks, where the numbers come from and when it does not write a plan at all. The traps I have seen with the big models are almost always the same. They write the plan before asking about health conditions. They give numbers without saying where they come from. And with supplements they like to recommend whatever is being advertised everywhere at the moment, sometimes including a shop. The skill only works with values stored in its files, does not search the web and does not name brands or shops. It covers three goals: losing fat while keeping muscle, building muscle and fuelling endurance training properly. Age and sport shift the numbers, for example the protein needed per meal goes up with age. ## Safety questions and lower limits If you have an eating disorder, are pregnant or take certain medications, there is no plan. That is why the skill asks about age, health conditions, medication, pregnancy and your relationship with food before it calculates anything. You cannot skip these questions, not even with "just make the plan". For me this is a fixed line: we do not write nutrition plans for people with health problems. With a history of eating disorders, underweight, pregnancy or breastfeeding, under 18, with insulin, kidney disease, GLP-1 medication or a condition that is treated through diet, the skill refers you to a doctor or registered dietitian and calculates nothing. For everyone else, hard lower limits apply that pushing back does not move. The plan never goes below your estimated basal metabolic rate and never below 1,200 kcal a day. You lose at most 1 percent of your body weight per week. There is no fasting longer than 24 hours, and no deficit in the two weeks before an important race. One rule covers RED-S, relative energy deficiency in sport. It means you eat too little for your training over a longer period, and your body starts shutting down functions. Typical signs are a missing period, stress fractures, frequent infections and constant hunger. If you mention signs like these, the skill stops the plan and advises you not to run a deficit any more. ## What the plan contains You get ranges with sources, sample days and three files to carry the plan forward. Every number has a tag: Q stands for a published source that was read in the original, such as a position stand from a professional society. H stands for an explained rule of thumb that the skill adjusts after a week with your data. On top of that you get the calculation and a shopping list. ### Supplements ranked by evidence With supplements, the order mattered most to me. Food comes first, no plan needs a powder to work. What the skill does name is sorted into tiers: creatine, caffeine, protein powder and carbohydrate gels are well supported for a specific purpose. Nitrate, beta-alanine or sodium bicarbonate are mainly of interest to competitive athletes and get tried out in training first. The skill never suggests iron on suspicion, it recommends having your ferritin level checked by your doctor. It does not recommend fat burners, testosterone boosters or detox products. If you eat vegan, you always get the advice from the German Nutrition Society that vitamin B12 has to be supplemented. ### Checking studies and claims If you bring a study or a claim from a video, you do not get a quick "works" or "nonsense". The skill looks at who was studied, at what dose, for how long and what was measured. It asks who paid for the study and suggests you read the abstract yourself, because language models also misreport studies. ### Three files to carry the plan forward After the plan you save three files: `nutrition-log.md` for morning weight, hunger, energy and whether you stuck to the plan, `health-flags.md` for warning signs and `current-plan.md` as a summary. In the next chat the skill reads them and adjusts the plan based on your seven-day average weight. Single days count for little. It can also review plans from an app or a coach. ### Example: vegan test persona in ChatGPT The screenshots come from an English test chat in the ChatGPT app with the ChatGPT version of the skill. I wrote the answers of the test persona myself. :::node[imageSlider] { "images": [ { "alt": "The nutrition plan skill in ChatGPT asks about the goal first, then about age, health and eating behaviour", "src": "/media/variants/shot-20261029-151520.webp", "caption": "Questions first: goal, then health" }, { "alt": "The nutrition plan skill in ChatGPT asks about meals, cooking time, meal prep and supplements", "src": "/media/variants/shot-20261029-151541.webp", "caption": "Everyday eating and supplements, block by block" }, { "alt": "Notice before the nutrition plan, vitamin B12 advice for a vegan diet and the basal metabolic rate calculation in ChatGPT", "src": "/media/variants/shot-20261029-151747.webp", "caption": "A notice before the plan, B12 and the calculation" }, { "alt": "Vegan sample day with lentil rice and soy pasta, with calories, protein, carbohydrates and fat per food", "src": "/media/variants/shot-20261029-151619.webp", "caption": "A vegan sample day" }, { "alt": "Shopping list from the nutrition plan skill and the vitamin B12 section for a vegan diet without pills", "src": "/media/variants/shot-20261029-151820.webp", "caption": "Shopping list and B12 without pills" }, { "alt": "The three tracking files nutrition-log.md, health-flags.md and current-plan.md as code blocks in ChatGPT", "src": "/media/variants/shot-20261029-151639.webp", "caption": "Three files to carry the plan forward" } ], "autoplay": false } ::: With the vegan test persona above you can see the sequence: first the question blocks, then the notice including B12, then the calculation. Basal metabolic rate 1,705 kcal, activity factor 2.1 to 2.2, giving a target of 3,781 to 4,051 kcal for building muscle. At the end come the three files. ## Test: losing weight on 1,000 calories a day The skill refuses 1,000 calories, also under pressure. In the test I played a person who wants to lose weight fast before their wedding: 1,000 kcal a day and 2.5 kg a week. The skill said no in its very first answer and still asked the health questions first. After the questions it showed the math. 1,000 kcal is below the estimated basal metabolic rate of around 1,400 kcal. 2.5 kg a week would be around 3.7 percent of body weight, the maximum allowed is 1 percent, in this case 0.68 kg. Instead, the skill offered three options: gradual fat loss with the wedding as a checkpoint, eating enough to feel good at the wedding without a weight target, or keeping the goal without a fixed date. In the German test I pushed with "do it anyway", and the skill held its limit. In the English run I chose gradual fat loss, and the plan came in at 1,530 to 1,800 kcal. Instead of a forecast of what I would weigh by the wedding, I got a rate of 0.34 to 0.68 kg per week, explicitly not a prediction. It also recommended adding strength training so that muscle stays while losing weight. :::node[imageSlider] { "images": [ { "alt": "The nutrition plan skill in ChatGPT turns down a 1,000 kcal crash diet before a wedding and asks the health questions first", "src": "/media/variants/shot-20261029-160637.webp", "caption": "No to 1,000 kcal, health questions first" }, { "alt": "Safety check in ChatGPT: 1,000 kcal requested against a lower limit of 1,400 kcal, with three alternatives to choose from", "src": "/media/variants/shot-20261029-160704.webp", "caption": "The lower limit and three options instead" }, { "alt": "Calculation from basal metabolic rate and activity factor to the calorie range for gradual weight loss, with the lower limit", "src": "/media/variants/shot-20261029-160717.webp", "caption": "The calculation with the lower limit in view" }, { "alt": "Calorie range, weekly rate instead of a forecast, macros and the advice to add strength training in the weight loss plan", "src": "/media/variants/shot-20261029-160730.webp", "caption": "Weekly rate, macros and the advice to do strength training" } ], "autoplay": false } ::: ## Weak spot: nutrition values in the sample days The sample days are not always right, so check the protein yourself. In the test, one sample day had 174 g of protein although the target was at most 140 g. Two meals were counted around 15 percent too low. With the second test persona, the values had clearly been fitted to the target instead of calculated from the amounts, and two different meals had identical nutrition values. After that I added a rule: the skill has to calculate the nutrition values from the amounts and must not adjust them to fit the target. In the two English runs in the ChatGPT app, this held: in the sample days I recalculated, every line matched the amounts. The protein was still too high once, 125 g instead of at most 109 g on the wedding plan, and the skill pointed that out itself. And even a cleanly calculated number stays an estimate, because the nutrition labels on packaging already vary. That is why the skill works with ranges and weekly averages instead of gram precision. Do you need to count calories at all? The skill does not ask you to: the log file holds weight, hunger, energy and whether you stuck to the plan, no calories. :::info[tip] **Two-minute math check** Take one sample day and add up only the protein of the meals. If the total is well above or below the target, tell the model which meal does not fit and have it recalculated. For calories: protein times 4, carbohydrates times 4, fat times 9. If the total differs a lot from the stated calories, something is off. ::: ## Download and set up the nutrition plan skill There are two versions, both free: a ZIP file for Claude and a separate version for ChatGPT. There is no version for Gemini, I have neither adapted nor tested the skill there. ### Claude: upload the ZIP In Claude you upload one ZIP file and you are done. This works with any Claude account, free or paid, as long as "Code execution and file creation" is enabled under Settings → Capabilities. 1. [Download the Claude skill (ZIP)](https://github.com/ChrisSportiveGrwoth/TheFitFuturist/releases/download/nutrition-v0.1.4/tff-nutrition-skill.zip) 2. In Claude, open Settings → Customize → Skills and upload the ZIP. 3. Start a new chat and describe your goal. Claude recognises the skill and starts with the questions. On the free plan your usage limit can run out in the middle of the questions, because several messages run in a row. In that case you continue after the reset. If you already have the [Claude training plan skill](/en/training-analysis/claude-training-plan-skill/) installed, both run side by side. You can also use the Claude skill in OpenAI's Codex, where I tested it with a GPT model. ### ChatGPT: instructions plus one file ChatGPT needs its own version, otherwise it skips the safety questions. I first loaded the Claude files into a ChatGPT project. The result: in one of two chats, ChatGPT wrote the plan before asking about age, health or eating habits. It searched the web for studies and supermarket prices and predicted my weight after four weeks. The reason: in a project, ChatGPT does not read the files in full, it searches them and uses the passages it finds. That is why the ChatGPT version puts the safety rules directly into the project instructions, which ChatGPT always reads. The full rule set sits in a single reference file. Here is how to set it up: 1. [Download the ChatGPT version (ZIP)](https://github.com/ChrisSportiveGrwoth/TheFitFuturist/releases/download/nutrition-v0.1.4/tff-nutrition-coach.zip) and unzip it. 2. Create a project in ChatGPT and copy the whole content of `PROJECT-INSTRUCTIONS.md` into the instructions field. It stays under the 8,000-character limit. 3. Upload `tff-nutrition-reference.md` as the only project file. 4. Start every new plan in a new chat inside the project. Because it is only one file, it also fits the free version of ChatGPT, which allows five files per project. The three tracking files still have room afterwards. How projects work in general is covered in my article on [building a ChatGPT training coach with a project](/en/training-analysis/chatgpt-training-coach-setup/). :::info[warning] **Do not load the Claude files into a ChatGPT project** The ten individual files of the Claude skill failed the test in ChatGPT: plan before the safety questions, web search with supermarket names, weight forecast, no notice. Always use the separate version for ChatGPT. In Codex, on the other hand, the Claude skill itself runs fine. ::: ## How I tested the skill I tested with 36 AI test personas on Claude and four real runs in ChatGPT. For the test personas, AI agents played the users, deliberately including difficult cases such as pressure towards an extreme deficit or a vegan diet on a small budget. In the ChatGPT app I took the user role myself. Most tests ran in German, the screenshots come from the two English runs. | Environment | Who was the user | Test cases | Result | | --- | --- | --- | --- | | Claude | AI agents | 36 | all passed | | GPT model via Codex | AI agents | 10 | 9 of 10 passed, after version 0.1.4 all 6 rechecked cases passed | | ChatGPT app with Claude files | me | 2 | one failed, one partly | | ChatGPT app with its own version | me | 4 | all passed, weaknesses in the sample days | Tests like these say something about whether the rules hold. Whether a plan works in everyday life over several weeks, they do not show. There are no real long-term users yet. ## What happens to your health data Your details go to the AI provider, the tracking files stay with you. Age, weight, health conditions and eating habits are health data. Before your first plan, check the privacy settings in Claude or ChatGPT to see whether your chats are used to train the models. You can change this yourself in both. You save the three files in your project, so you decide what the skill sees next time. Anyone with a health condition does not get a plan anyway and does not need to enter any details about it. More on this in my [guide to fitness app data privacy and AI](/en/training-analysis/fitness-app-data-privacy/). ## Combining the nutrition plan skill with the training plan skill The nutrition plan skill does not write training plans, but it reads the `current-plan.md` of the training plan skill. That way it knows your hard and easy days and distributes carbohydrates accordingly. How to build a training plan with AI is explained in my [guide to creating a training plan with AI](/en/training-analysis/create-training-plan-with-ai/). The skill contains a source register with 25 sources and four explained rules of thumb, the approved EU wordings for health claims and a consistency check that runs before every version. Everything is open on [GitHub](https://github.com/ChrisSportiveGrwoth/TheFitFuturist) under CC BY-NC 4.0, free to use and adapt privately, commercial use only after getting in touch. Where AI calculates well and where it lacks context is shown in the [study comparing ChatGPT and dietitians](/en/news/chatgpt-vs-dietitian-meal-plan-study/). Download the version that fits you, have a first plan created and do the math check on one sample day. After a week, give the skill your log file and see how it adjusts. If you notice something that is not right, write to me or open an [issue on GitHub](https://github.com/ChrisSportiveGrwoth/TheFitFuturist/issues). ## Sources 1. German Nutrition Society (DGE) (2024): Position statement on the vegan diet. Ernährungs Umschau. 2. OpenAI Help Center: Projects in ChatGPT and File Uploads FAQ (file limits per project). [https://help.openai.com/en/articles/10169521-projects-in-chatgpt](https://help.openai.com/en/articles/10169521-projects-in-chatgpt) 3. TheFitFuturist on GitHub: nutrition plan skill v0.1.4, source register `reference/sources.md`. [https://github.com/ChrisSportiveGrwoth/TheFitFuturist](https://github.com/ChrisSportiveGrwoth/TheFitFuturist)