Claude Workout Plan (and Gemini): Prompts and Limits

By the Trunk teamUpdated September 7, 202614 min readPrices verified September 7, 2026

Claude and Gemini will both write you a training program, and the structure is usually good enough to run. The catch is memory: a Claude workout plan is built from what you typed into the chat, not from what you actually lifted, so by week four the model has no idea whether you hit the loads it prescribed. Below are two prompt templates, the one setting Claude needs before it can open a link, and every route we could verify for feeding your real logged sets back in. Disclosure: this blog is published by Trunk, one of the apps compared below.

What a Claude workout plan gets right, and where it stops

Both assistants handle the structural parts of programming well: splitting a week, spreading sets across muscle groups, putting compounds before accessories, writing a deload. Neither handles the part that decides whether you get stronger, which is reacting to what happened on the bar on Tuesday.

Claude needs web search switched on, per chat

Anthropic's help center describes turning it on from the "+" button in the chat composer, and says that once it is enabled, Claude "can retrieve full content directly from specific URLs you provide" (source). Skip that step and Claude will tell you it cannot open your link.

That article documents the toggle on the web app and says nothing about the iOS app's menu, so if Web search is not there on your phone, paste the prompt at claude.ai in a browser instead. The same page warns that a long linked document is pulled into the context window whole and counts against your daily usage, which bites on the free plan. Claude Free is $0 and lists web search among its features, verified 2026-09-07 (source).

Gemini works, but Google does not document it

Google's help center says Gemini "can always use public information from Google Search" (source), and the free plan is $0 as of 2026/09/07 (source). No Google page states that Gemini will open a URL you paste into a chat. It does in our own testing — we build Trunk and we have run the flow — but there is no vendor document to point at, so treat that as tested behavior rather than a guarantee.

If you are running Gemini as a workout planner, the other free-tier limit is context size: 32k tokens on free, 128k on AI Plus, 1M on AI Pro and Ultra (source). A long lifting history outgrows the free window, so narrow what you send.

Prompt template 1: review the program you are already running

Start here rather than with a blank slate. A review forces the model to read your data first, which also confirms the link went through.

Act as a strength coach reviewing an existing program. Do not rewrite it yet.

Here is my training log: [paste your data link, or attach your export]
Open it now. Before anything else, list the programs, days and exercises
you actually found, so I can confirm you read the real file.

Then answer in this order:
1. Weekly hard sets per muscle group, counted from the log.
2. Which lifts have stalled, with the dates and loads behind that call.
3. Anything programmed but rarely logged.
4. The three changes you would make, ranked, each with its reason.
5. What in this log would make you send me to a physio instead of
   changing the program.

Context: I train [N] days a week, [minutes] per session.
Equipment: [list]. Injuries or limitations: [list, or "none"].
If the log does not support a claim, say so rather than guessing.

Line 1 is worth reading twice. Weekly set counts per muscle group are tedious by hand, trivial for a model that can see every logged set, and usually where a self-written program turns out lopsided.

Prompt template 2: build a new training block

Build me a [N]-week training block.

Read my log first: [paste your data link, or attach your export]
Base starting loads on what I have actually lifted, not on estimates.

Goal: [hypertrophy / strength / both]
Days per week: [N]   Session length: [minutes]
Equipment available: [list]
Lifts to prioritize: [list]
Injuries or limitations: [list, or "none"]

Output rules:
- One line per exercise: sets x reps @ load or RIR.
- State the progression rule explicitly (add load, add reps, or hit an
  RIR target) and say what triggers it.
- Include a deload week and say exactly what changes in it.
- Give each training day a name so I can log it as its own day.
- No supplements, no diet plan, no medical advice.

Those output rules do real work. Without them both models drift into "increase weight when it feels easy", which is not a progression rule. Name the trigger and you get something you can run without re-reading the chat.

Treat the first draft as a draft. One reviewer comparing ChatGPT and Claude for personal trainers flagged ChatGPT drifting into medical territory without flagging that a referral was needed (source). If you are carrying an injury, coming back from surgery, or managing a condition, have the plan checked by someone qualified before you run it.

Where a Claude workout plan breaks: it cannot see what you lifted

The obvious fix looks like connecting Claude to Apple Health. For a lifter it does not work. Anthropic's own iOS support page lists the Apple Health workout data Claude can read as "Type, duration, distance, heart rate, calories burned" (source). Sets, reps and weight are not on that list. Your top set of 315 arrives as a workout type, a duration and a calorie estimate, with nothing about the bar inside it.

That page adds three limits: Pro and Max only, US users only, read-only. The Android equivalent needs Health Connect and Android 14 or later, and does not work under a work profile (source).

Gemini's official route is Google Health Coach, behind Google Health Premium at $9.99 per month or $99 per year and included with Google AI Pro and Ultra. Google says it went globally available on May 19, 2026 and connects to Fitbit devices, Pixel Watch, Health Connect and Apple HealthKit, verified 2026-09-07 (source). Whether per-set load and reps survive that trip we could not verify. Either way the health stores are not the path — you have to hand the model the log itself.

How to hand Claude or Gemini your real logged sets

HevyGPT: mature, and ChatGPT only

An official custom GPT connected to Hevy: it analyzes your training history and writes generated routines back into your profile. Hevy states plainly that you do not need Hevy Pro and that a free ChatGPT account works, but free Hevy accounts are capped at four routines, so a bigger plan cannot be imported without paying (source). No Claude version exists, and no Gemini version. Hevy is also the grown-up here: iOS, Android, Apple Watch, Wear OS and web, 4.9 stars from 91K US App Store ratings, verified 2026-09-07 (source).

Hevy's send-to-Claude button

Hevy added "Send workouts to ChatGPT & Claude" from the finish-workout screen in iOS 3.1.12, released September 3, 2026. From the release notes it appears to be a one-way share of a finished workout, but Hevy has published no help article, so the mechanism and whether it needs Pro are unknown (source).

Liftin' MCP: ChatGPT and Claude, in beta

Liftin' announced an MCP server on July 22, 2026 as "full integration with ChatGPT or Claude" for analyzing history, changing routines and creating a program (source). Gemini is not named, setup lives only in a Reddit post, and nothing official says whether the beta needs Liftin' Unlimited — $2.99/mo, $24.99/yr or $99.99 one-time as of 2026/09/07, above a free tier of five workouts a month (source).

Arvo's MCP server: free and open source

The cheapest route here, and worth trying before any paid one. Arvo describes arvo-mcp as "completely free and open source", exposing 29 tools and full read and write access to workouts, PRs, exercises and splits from Claude (source). You need an Arvo account and an MCP client you are willing to configure.

Strong: export only

Strong exports CSV and imports nothing. Upload the file to Claude or Gemini and analysis works fine; getting a plan back in is manual, because Strong's help center says exported files cannot be imported back (source), and free accounts are capped at three custom routines (source). Strong has no AI at all — but it is the fastest logger in the category, has the best standalone Apple Watch app of any lifting tracker, and carries 4.9 from 109K US App Store ratings (source).

Ellim: import only

Ellim goes the other direction. Its free tier turns a pasted plan, or a photo of one, into a trackable routine with no routine cap, so a Claude or Gemini program lands cleanly. No outside AI reads your history. Ellim's own AI costs $17.99 per month or $99.99 per year with no one-time option, as of 2026/09/07 (source). iPhone-only.

Trunk: a plain URL

Generate a key, copy the ready-made prompt with its private data link, paste it into any assistant. Nothing to install, no developer key, no MCP server. Claude still needs its web-search toggle on first, and Trunk's prompt carries an explicit instruction block because ChatGPT sometimes answers from memory instead of opening the page.

Features and prices verified 2026-09-07. Figures are USD from the US App Store and vendor pages; prices vary by region and may change. The Claude and Gemini cells in the Trunk row are our own testing, not vendor-documented.

Anyone living in ChatGPT and Hevy has no reason to move, and anyone comfortable configuring an MCP client should try Arvo first, because it costs nothing. Trunk's argument is narrower: Claude and Gemini users who want no setup beyond copying a prompt, and anyone who would rather not add a monthly line item. On the ChatGPT side, see the ChatGPT connection guide and the HevyGPT and Trunk comparison.

The seven-step flow, with Claude or Gemini

1. In Trunk, open Settings, then Trunk+, then Connect AI.
2. Tap Generate Key. Trunk creates a random token and builds a prompt around it.
3. Tap Copy Prompt. The prompt contains a private data URL plus instructions telling the assistant to open it.
4. In Claude, turn on Web search from the "+" button first. Gemini has no equivalent toggle to find.
5. Paste the prompt. The assistant reads every profile, session and exercise, with reps, weight, volume, per-set notes, timestamps and plate configuration.
6. Want a block written? Say so, and the reply arrives as a single import link. Open it, read what the confirmation page says it will create, then tap Add to Trunk.
7. Hit Revoke Key when you are done. Generating a new key also invalidates the old one.

Query parameters on the same URL narrow what the assistant reads — ?exercise=bench, ?profile=Name, ?session=Push, or ?last=5 for the five most recent sets per exercise. Given Gemini's 32k free context and Claude's whole-page fetch, narrow the request when your history is long. We have not measured how large a Trunk page gets for a multi-year log.

The write side is deliberately narrow: create a program, add days, add exercises, up to 20 days and 30 exercises per day, and it will not overwrite a profile you already have. It never writes sets, reps or weights — you log those in the gym, so the record stays real.

One warning before you paste anything: that URL is the key itself, not a link protected by one. Anyone who has it can read your training data until you revoke it. Chat transcripts get screenshotted and shared more often than people expect, so if a conversation containing the link ever leaves your device, generate a new key — doing so kills the old one immediately.

What a year of this costs

Prices verified 2026-09-07, in USD, from the US App Store and vendor pages; they vary by region and may change. Hevy Pro is not an AI subscription — it appears because it is what the HevyGPT route costs once you outgrow the free four-routine cap.

Google charges $9.99 a month, or $99 a year, for its AI coach (source). Trunk charges $9.99 once, because it runs no model of its own and has no per-user model bill to recover — you bring the assistant you already have. That is why a one-time price works here and not there. It is not the cheapest route anywhere, though: Arvo's server is free, Hevy sells a $74.99 lifetime (source), and Strong sells a $99.99 Forever unlock as listed, with a $79.99 SKU also on the page (source).

Who should pick which

ChatGPT user already on Hevy: stay put. HevyGPT is free, bidirectional and needs no Pro, and Hevy brings Android, Apple Watch, web and 91K US App Store ratings.

You want this to cost nothing: Arvo's arvo-mcp — free, open source, read and write from Claude — if you will keep your log in Arvo and configure an MCP client.

Android, an Apple Watch app, or an exercise library with demos: Hevy, Strong or JEFIT. Trunk is iOS only, needs iOS 26.2 or later, and ships no Apple Watch app, no Android app and no exercise library.

Claude or Gemini user who wants the model reading real sets, with no subscription and no MCP setup: Trunk+ at $9.99 once, on iOS. The trade is real — Trunk is new, with 1 App Store rating against Hevy's 91K and Strong's 109K, and no CSV export button. What it does is the plain-URL connection, in any assistant, with no caps.

Staying on Strong: export the CSV, upload it, retype the plan by hand inside the free three-routine cap. It costs nothing and works today. If you would rather the plan arrive as a tappable link, that is what importing an AI plan covers.