How to Make Claude Your Personal Trainer

Claude works as a training partner that writes programmes, reads your log and answers questions, and it has one structural advantage over other assistants for the nutrition half: it can call a food database instead of recalling one. Here is the setup, the prompts, and the boundary.

Disclosure first: we build a nutrition database, not a fitness product. The training guidance below is general and widely published, and we say so rather than implying expertise we do not have. The food half is ours.

What can Claude do as a personal trainer?

The same things any capable assistant can, plus one thing most cannot.

It writes programmes well. Splits, exercise selection, set and rep schemes, adapted to your days, equipment and constraints. Standard training templates are heavily documented.

It reads a log properly. Give it six weeks of sessions and ask what stalled, and you get a useful answer. Text analysis is what these systems are genuinely reliable at, and spotting a lift that has not moved is something humans do badly.

It can look food up. In Claude Code or Cursor you can attach a nutrition tool over the Model Context Protocol, so the food side stops being recalled and starts being retrieved. That is the difference from a plain chat window, and it matters because the nutrition half of a training plan is where checkable errors accumulate.

It cannot see you. Same as everything else. No form correction, no technique teaching, no noticing that you are under-recovered.

Setting up the memory

The failure mode of any assistant as a trainer is starting fresh every session. Fix that with a file the client reads automatically. In Claude Code, a markdown file in the working directory works:

# Training profile
- 34, male, 82 kg, 180 cm. Two years lifting.
- 4 sessions a week, 45 minutes, full gym.
- Left shoulder dislikes overhead pressing.
- Current bests: squat 120x5, bench 85x5, deadlift 150x3.

## Rules
- Never prescribe a specific weight. Give rep ranges with an RPE target.
- Append to the session log. Never rewrite earlier entries.
- If I report pain, tell me to see a professional. Do not program around it.
- One change per review. Do not rewrite the programme unprompted.

## Session log

Those four rules do more for output quality than any prompt, because they apply to every reply rather than one. The last one is the most underrated: assistants will happily hand you a brand new programme whenever you ask a question, and most training results come from repeating something adequate long enough to progress on it.

The prompts

Writing the programme

Write a 4-day upper/lower split from my profile. No specific weights,
rep ranges with RPE targets. State the progression rule and when it
will stop working. Flag exercises that commonly cause trouble and give
one swap each.

After a session

Session: squat 100x5x3 RPE 8, bench 75x5x3 RPE 9, rows 60x8x3. Append
to the log. What changes next session for each lift, and flag anything
that has not moved in three sessions.

Reviewing a block

Read the last six weeks of the log. Which lifts progressed, which
stalled, and what single change would you make? Do not rewrite the
programme.

The nutrition side

I trained today. Log this meal and tell me where I am against 175 g
protein: 200 g chicken thigh, 250 g cooked rice, 200 g broccoli,
15 g olive oil. Search the catalog for each and scale it.

That last one is where the database earns its place. Without it the protein figure is recalled. With it you can ask which catalog row produced each number.

Connecting the nutrition tool

Claude Code:

claude mcp add --transport http calorie-api \
  https://calorieapiadmin.com/mcp \
  --header "X-API-Key: YOUR_KEY"

The key comes from the dashboard after starting a plan on MCP pricing. Claude.ai, Claude Desktop and Claude mobile cannot connect to this server, because their connector flow expects browser sign-in and that is not enabled, so use Claude Code or Cursor. The walkthrough is in connect a remote MCP server, the reference is in the docs, and the plan overview is on the MCP page. Eating for training is in tracking a lean bulk.

A worked first month

The setup is the easy part. This is where people fall over.

Week one. Write the profile. Get the programme. Change nothing and record every session, including the ones that went badly, because a log with holes cannot tell you anything in week six.

Week two. Same programme. Start using the after-session prompt so the log builds. The strongest urge in the first fortnight is to ask for a different plan; resist it.

Week three. First review. Ask what moved and what did not. Expect at least one stall. A stall in week three is normal and not evidence the programme is wrong.

Week four. Make one change. Not five. The value of a log is being able to attribute an outcome to a change, and altering several things at once destroys that.

One change per review is what separates training from collecting plans, and it is the hardest discipline here because generating a new programme is instant and free while repeating one for six weeks feels like doing nothing.

How do you know it is working?

Decide the criteria before you begin, because judging afterwards is where wishful thinking lives.

For strength, the signal is whether the main lifts move over six to eight weeks, allowing for a bad week. For body composition it is slower and noisier, and measurements plus photographs across months beat a mirror on any given day.

Not signals: how hard a session felt, how sore you are afterwards, or how impressive the programme looks written down. Soreness in particular is a poor proxy for anything useful, and chasing it is why people change programmes constantly and progress on none of them.

Write those criteria into the profile file so reviews are run against them rather than against how you happened to feel that week. An assistant reading a log is genuinely good at this, because it has no memory of wanting the answer to be yes.

Three things never to take from it

A starting weight. It is guessing about a body it cannot see. Take the structure, set loads from your own recent performance.

Advice about training through pain. Persistent pain is a physiotherapist question. The standing rule in the profile file is what stops it offering confident modifications.

Supplement recommendations. Evidence is contested, regulation differs by country, and interactions with medication are a medical question. Ask a pharmacist.

A prediction of what you will achieve. Any figure for how much you will lift or gain in twelve weeks is invented. Training response varies enormously between individuals, and a confident projection is the clearest sign you are reading generated text rather than reasoning.

Nutrition figures without a lookup. The reason this article recommends attaching the database. A protein number quoted without a catalog row behind it is a guess wearing a decimal point, and across a training block those guesses compound in one direction.

Claude Code or the Claude app?

A practical question, since the answer changes what you can do.

Claude Code or Cursor. Can attach the nutrition tool. Reads and writes files, so the profile and session log work automatically. This is the full setup and it is a desk workflow.

Claude.ai, Desktop or mobile. Good for programme writing, explanation and reviewing a log you paste in. Cannot attach this server, because browser sign-in for remote connectors is not enabled here. Convenient in the gym, which is where you actually are.

A reasonable split is to use the app for asking questions and recording sessions on your phone, and the desk setup for the weekly review and anything involving food figures. The log lives in a file either way; you just paste it when you are on mobile.

What it cannot do, stated plainly

It cannot see your form. No amount of describing a lift replaces someone watching it.

It has no view of recovery. Sleep, stress and work load determine what you can handle, and a set-and-rep log records none of it. You have to supply that context and it will not ask.

It has no accountability. Nobody is expecting you on Tuesday. If that is why a trainer worked for you before, this does not replace it.

It is agreeable. Ask whether your programme is good and it will find reasons it is. Ask what is wrong with it and you get something useful. Phrase reviews adversarially, every time.

It will not tell you to stop. Deloads, rest weeks and doing less are part of training, and an assistant will rarely volunteer them because they are not interesting outputs. Ask directly.

Where the boundary is

A training plan cannot be verified against data. Whether four sets of eight suits you is judgement, built from watching many people, and no assistant has that.

Food composition is different. How much protein is in 180 g of cooked chicken thigh has a right answer in a table, which is why the nutrition half is the half worth making rigorous. Getting that half right is a data problem you can solve. The training half is a judgement problem you cannot, and the honest move is to use published programming principles and a qualified human where it matters.

If you are new to lifting, returning from injury, pregnant, or managing a condition, have a professional review the programme. Technique under load is the largest injury risk in a gym and the least suited to written instruction.

What we have not measured

No training outcomes, no comparison against coached programming, no effectiveness claims. We are a nutrition data company. We have also not published an accuracy benchmark for our own food data.

What the post contains instead is the mechanism: which part of the workflow produces which number, which parts are judgement and which are lookup, and where the boundary between them sits. Those you can verify yourself in an afternoon, which is more durable than a claim you would have to take on trust.

Related reading before you start

If you are setting this up from scratch, the order that wastes least time is: attach the nutrition tool first and confirm it returns real rows, then write the profile file, then get a programme, then start logging sessions. Doing it in the reverse order means rebuilding the log once you realise the figures were not coming from anywhere.

Related

• How to make ChatGPT your personal trainer
• AI personal trainer: what you get and what you do not
• Tracking a lean bulk in Claude
• How to make Claude your personal dietitian