How to use ChatGPT to plan your marathon training

You can ask ChatGPT to build you a marathon training plan in thirty seconds. The output looks impressive. The plan has weeks, paces, intervals, taper. A real coach reviewing it will mostly nod.

The problem is what happens next. The plan sits in a chat window. You screenshot it. You forget half of it. By week five you're running by feel and pretending the plan exists.

This post is about the workflow that actually works in 2026: how to prompt ChatGPT (or Claude, or Gemini) for a usable marathon plan, what to verify, and how to get it onto a watch where you'll actually follow it.

The prompt that produces a usable plan

Generic prompts ("give me a marathon training plan") produce generic results. Pace zones missing, mileage random, no taper. Useless.

The prompt that works is specific. Give the model your context:

"I'm training for a marathon on [date]. My target time is [HH:MM] (so [pace]/km goal pace). My current weekly mileage is [X] km, my recent half marathon was [HH:MM] in [Month Year], my long run is comfortably [X] km. I have [N] days a week to train. I prefer [methodology - Pfitzinger / Daniels / Hanson / Higdon / flexible]. Build me a [12 / 14 / 16 / 18]-week plan, day by day. Use km not miles. Include pace targets for each quality workout based on my goal pace. Show me the plan as a markdown list with one week per heading."

This prompt does three things that matter:

1. Anchors the model in your current fitness so it doesn't prescribe 60 km weeks to a 30 km runner
2. Specifies a methodology so the model doesn't blend everything into mush
3. Forces a structured output format so you can parse and re-use the plan later

You'll get a plan that looks like the ones in our first post. Week-by-week, day-by-day, with paces and intervals.

What to verify before you trust the plan

ChatGPT plans are not malicious but they make systematic mistakes. Check these before week one:

Mileage progression: a sane plan grows weekly mileage by no more than 10% per week, with a recovery week every 3-4 weeks where mileage drops 20-30%. If the model gives you a steady ramp with no recovery weeks, ask for the recovery weeks.

Long run percentage: your longest long run should be roughly 25-35% of your weekly mileage at peak. If it's 50%+, the plan is back-loaded and risky. If it's less than 20%, you're not going to have the endurance for race day.

Pace targets: if all your interval and tempo paces come out at "5:00/km" and "4:30/km" no matter what goal time you gave, the model is hallucinating round numbers. Ask it to derive paces from a recent race performance using the VDOT or McMillan calculator method.

Taper length and depth: a marathon taper is typically 2-3 weeks, with mileage dropping to ~60% of peak in the final week. If the model drops you straight from 70 km peak to race week with no in-between, ask for a proper taper.

Race-pace exposure: somewhere in the plan, you should have several long runs with embedded marathon-pace segments (e.g. "20 km with last 8 km at MP"). If the plan is all easy runs and intervals with no marathon-pace work, your legs will be shocked on race day.

If the plan fails any of these checks, ask the model to fix that specific thing. "Add recovery weeks at weeks 4, 8, and 12" is a precise instruction it will execute.

What ChatGPT can't do (yet)

A 25-time marathoner who reviewed a ChatGPT marathon plan recently called the structure "spot on." A coach with 30 years of experience will tell you the same plan misses things only a coach can:

• It can't tell you that you sound exhausted in week 6 and need to back off
• It can't adapt the plan based on how a workout actually felt vs. how it was supposed to feel
• It won't tell you when you're being stupid
• It has no idea your knee has been twinging for a week

For most self-coached amateur athletes, that's fine. You know your body. You can call the audibles yourself. You're using the AI for the structure, not the judgement. And structure costs $30-300/month from a human coach, but $0 from ChatGPT.

The trade-off is honest, and it's the right trade-off for a lot of people. Just don't pretend you're getting coaching. You're getting structured suggestions that you have to actually use your brain to follow.

The workflow that actually gets the plan to your watch

This is where most people lose the plan. The model produced a great document. The athlete copies it to Notes. The Notes app does nothing useful with it.

The workflow that actually keeps you on the plan looks like this:

1. Generate the plan with the specific prompt above
2. Verify it against the five checks above
3. Paste the markdown into a tool that parses it into structured workouts
4. See it on a calendar so you know what's coming this week and next
5. Push each day's workout to your watch so when you start a run, the watch already knows the intervals
6. Track adherence so you see whether you're following the plan or drifting

Today, step 3-6 is what's broken for most people. There's no clean path from "markdown plan in ChatGPT" to "structured workout on my Apple Watch this morning."

That's the gap we're closing with Stopa. Pre-launch, iOS-first. If this is the workflow you want, join the waitlist.

A note on AI plan safety

There's been a wave of stories about AI-generated training plans being too aggressive and causing injuries. Most of the cases involve subscription apps that algorithmically generate plans without context (the user's actual mileage, their injury history, their age). Generic LLMs like ChatGPT are not better or worse at this - they're as safe as the prompt and the verification step you apply.

The single most important safety practice with any AI-generated plan: never trust a weekly mileage jump greater than 10%. If you've been running 40 km/week and the plan starts you at 55 km/week, the plan is wrong. Have the model rebuild it with a 10% rule applied. Or just throw away the first three weeks and start where you actually are.

You're the safety check. The AI is a faster draft, not a coach.

Try this today

Open ChatGPT (or Claude, or Gemini). Paste in the prompt template at the top of this post, filled in with your details. Get a plan. Run it through the five-check verification. Then think about how you're going to actually use it for the next 16 weeks.

If the answer to "how" includes "screenshot, print, hope I remember" - you've got the same problem everyone has. We're building the fix. Coming soon to iOS.

Stopa takes any training plan from ChatGPT, Claude, your coach's Excel, or a PDF, and pushes the workouts to your Apple Watch via WorkoutKit. Pre-launch waitlist at stopa.app.