Asking an AI model to write a training plan is easy. Getting a useful plan, keeping it attached to what you actually did, and moving the workouts onto your watch is the harder part.
The workflow I find more convincing is a small weekly loop:
This is not a replacement for a coach, and an activity file is not a medical history. It is a practical way to reduce the administrative work around a plan while keeping the decisions visible.
A model knows nothing about your training unless you give it something useful. For a first review, 8 to 12 weeks is usually enough to describe the shape of recent training without dropping an entire account archive into the conversation.
Useful details include:
You rarely need to upload every GPS point. Location traces are sensitive and add little to most planning conversations. A compact activity summary is easier to inspect and easier for the model to reason about.
Strava now provides an official MCP connector for Claude. It gives eligible Strava subscribers read-only access to live activity history and training information through an OAuth connection. Strava says it can work with activity history, fitness trends, readiness, goal planning, cross-sport activity and gear.
In Claude, open Customize > Connectors, find Strava, choose Connect, and approve access in Strava. The rollout began in June 2026, so it may not yet appear for every subscriber.
The connector cannot upload or edit activities. You can revoke it later from Strava > Settings > My Apps. Strava documents the current availability and setup in its Strava MCP Connector guide.
Start by asking for analysis rather than a plan:
Review my last 8 to 12 weeks. Summarise weekly volume, frequency, long sessions, harder sessions, pace or power trends, easy-day intensity, cross-training, and meaningful changes in the last 7 and 28 days. Do not create a plan yet. Tell me what context is missing.
That last sentence matters. A confident answer based on incomplete information is still incomplete.
Strava lets you request a bulk account export and export individual activities as original, TCX or GPX files. Its export guide explains each route.
For this job, begin with the activity summary in the account export. Pull in a detailed activity file only when a particular session needs closer inspection.
Garmin Connect also lets you export the activity list as CSV, export individual activities, export selected reports as CSV, or request a full account archive. Garmin's data export instructions cover the current options.
The activity-list CSV is the sensible starting point. A full account export can contain considerably more health and location data than the model needs. Review any file before uploading it, and remove columns that are irrelevant to the question.
Training data can show that Tuesday's run was slower. It cannot reliably tell whether that happened because of fatigue, hills, heat, illness, a bad night of sleep or stopping to help somebody find their dog.
Add a short check-in covering:
This is also where the model should stop. Pain, illness and uncertainty about what is safe belong with an appropriate clinician or qualified coach, not a more elaborate prompt.
Workout Writer has a personalised AI Prompt Guide under its Help menu. It explains the simple text format the app expects and includes the names of your configured tags.
Those tags are the useful bit. If Easy, Tempo and Threshold already carry your preferred pace, heart-rate or power targets in Workout Writer, the model can use the names directly:
15 min Easy
3 x 10 min Tempo, 3 min Easy
10 min EasyYou do not need to repeat every target in the planning prompt. Workout Writer applies the settings you already maintain.
If you use a local agent on your Mac, it can fetch the same guide through Workout Writer's CLI bridge. Practical examples are in From AI Conversation to Structured Workout.
Keep a small reusable planning note with three sections:
Then give the model some boundaries:
Preserve the overall plan. Change only sessions affected by new evidence. Explain each change. Be conservative when the evidence is ambiguous. Return the next seven days, with each workout in a separate block using the attached Workout Writer guide. Challenge choices that do not fit my recent training rather than simply agreeing with me.
The model does not need to generate the entire season every Sunday. Plans become rather twitchy when one poor run causes the next three months to be redesigned.
Check the proposed week for ordinary mistakes:
AI should propose. You should approve.
There are two practical routes.
On iPhone or iPad, copy or select one approved workout and send it to Workout Writer through the Share Sheet. You can also paste the text directly into the app. Check the parsed steps, make any changes, then schedule it for Apple Watch or Garmin.
On a Mac, a local AI agent can use Workout Writer's CLI bridge to read the personalised guide, inspect existing workouts, create the approved workout, and schedule it to Garmin. Write and scheduling tools stay behind separate permissions. See From AI Conversation to Structured Workout for practical examples, or the Agent Access guide when you are ready to set it up.
Agent Access scheduling is currently Garmin-only. For Apple Watch, create and review the workout in Workout Writer, then use the app's normal Apple Watch flow.
Apple's Use Model action in Shortcuts can send input to an on-device Apple Intelligence model, Private Cloud Compute, or ChatGPT. Its Follow Up option lets you refine the response before it continues to the next action. Availability depends on the device, language and region. Apple explains the current options in Use Apple Intelligence in Shortcuts.
A useful shortcut could:
The Strava connector and Shortcuts are separate pieces. Strava currently documents its MCP for Claude, not as a Shortcuts data source. The shortcut is most useful for collecting your subjective check-in and handling the approved result.
At the end of the week, compare planned and completed sessions. Note what was done, skipped, shortened, extended or moved, plus a sentence about why. Refresh the rolling summary and ask the model to adjust only the next seven days.
That is less dramatic than generating a twelve-week plan in one prompt. It is also closer to how training works. The useful plan is the one that still knows what happened last Tuesday.