Using AI to Write Your Own Training Plan

An AI model can write you a sensible training plan, provided you give it the right context and know how to read what comes back.

• A share of the generic plans sold online comes straight out of a chatbot. If that’s the method, you may as well learn to run it yourself.
• Context comes before the request: three months of history in numbers, physical constraints framed as operating limits, the days you genuinely have, a goal with a date on it.
• Standing custom instructions keep the model inside sensible guardrails without you repeating them in every conversation.
• Checking the output decides everything: progression, cutback weeks, intensity distribution, paces that line up with a time you have actually run.

Twelve weeks in a PDF, a price tag, and the feeling that the plan could belong to anyone.

The Plan You Paid For Was Written by a Chatbot

A share of the plans sold online start life inside a generative AI model and reach you through someone who added a price and not much else. Thirty seconds is enough to spot them: the same architecture for everybody, the same four sessions with the numbers nudged around. They read like a horoscope. They feel written for you because they say nothing specific about anyone.

That’s a charge against the practice, and it leaves the profession alone. A coach who works through your actual numbers and answers you when the week falls apart is doing something else, and charges for it with reason. But if the file you bought came out of a chatbot, you may as well learn to get a good one out of it yourself. Almost all of the work sits in two places: what you write before you ask, and what you check afterward.

Build the Context Before You Ask for Anything

The difference between a useful plan and a list of workouts comes down to how much true information you put in. The wording of the prompt matters far less.

Start with the last three months, in numbers: weekly mileage, how many runs, the longest run of each month, the pace you hold when you run easy, your most recent time over a measured distance. Then the physical constraints, written as operating limits rather than symptoms to be interpreted: I’ve had Achilles trouble in the past, so no downhill repeats; I can’t run more than four sessions a week.
A language model has no standing to tell you what’s wrong when something hurts, and asking it remains the fastest route to a wrong answer delivered well.

Then your real availability, day by day, with the minutes you actually have: which days work has locked down, where the one long window of the week sits. If you claim five sessions because you’d like to run five, the plan is already wrong by the second Thursday. Last, the goal with its date: distance, race day, the time you have in mind. Without a date, the model builds a calendar of workouts instead of a progression.

The request, by then, more or less writes itself: a week-by-week table, total volume at the end of every row, paces in minutes per mile derived from the time you gave it, and a list of the assumptions it made wherever data was missing. That last item earns its place, because it shows you where the model filled a hole on your behalf. Inventing information when you leave a gap is endemic to these systems.

The Custom Instructions That Change What You Get Back

Nearly every assistant has a section of instructions that stay live across every conversation, and filling it in saves you from restating the premises in each new chat. Four guardrails cover most of the trouble.

The first covers volume growth: ask that it never exceed ten percent a week, and that every fourth week comes down. It’s a cautious convention more than a proven number, and it works well as a guardrail on a tool that, left alone, climbs in a straight line to the last row. Second: when a piece of information is missing, it asks for it instead of estimating. Third: every plan arrives with weekly totals, so the inconsistencies stay visible. The fourth covers pain, and it’s worth spelling out in full: no reading of symptoms, no guesses at a diagnosis, refer you to a professional.

I’d add a fifth, less technical: no encouragement. The lines spent telling you what a great question that was could have held something useful.

Checking What the Model Hands Back

Add up the mileage week by week and look at the curve before you read a single workout. This is where these tools fail most reliably: the progression rises in a straight line to race day, without the dips where adaptation settles, and the taper, when it exists at all, runs too short.

Then count the hard sessions in each week. More than half and the plan tilts the wrong way, because the distribution that holds up over months keeps most of the volume at low intensity, the way polarized training works. Look at where they land, too: two hard days back to back, or the long run the morning after intervals, tell you it filled the open days without thinking about how to spread your mileage across the week.

Paces deserve a check of their own. The model sets them from what you told it, so an optimistic estimate comes back to you multiplied by twelve weeks. Take a time you have actually run, training efforts included, and check the proposed paces against it: a 10K test — 6.2 miles — is still the cheapest way to calibrate the whole table against something real.

Then ask questions. Ask why the long run sits on Saturday rather than Sunday, what changes if the flu takes out a week, which sessions to drop if this month gives you three days instead of four. A plan built on reasoning survives those questions and tells you what you give up in each scenario. An assembled plan rewrites everything from scratch every time, or just reassures you and agrees, with the confidence of someone giving directions through a city they have never visited: the street names sound plausible, and the route goes nowhere near where they said.