A fair amount of traffic to this site arrives from ChatGPT, which tells you something about how people plan training in 2026. The chatbot has quietly become the first stop for "write me a programme", and the honest verdict on what comes back is: better than most of the internet expected.
This is a field guide to that workflow. What the AI plan does well, the four places it stops helping, and how to carry it from the chat window to the squat rack.
What ChatGPT gets right
Programme design for a standard goal is mostly settled knowledge. Hit each muscle with enough hard sets per week, spread them across the days you have, pick movements that fit your equipment, add load or reps over time. That consensus is all over the training literature the model was trained on, so when you ask for a four-day hypertrophy block, the output usually resembles what a sensible coach would sketch: compound lifts first, accessories after, volume in a defensible range, a deload mentioned somewhere.
It is also patient in a way templates never were. You can tell it you have dumbbells and a bench and nothing else, that your shoulder dislikes overhead pressing, that Thursdays are impossible, and it will redraft without complaint. That interrogation loop (ask, adjust, ask again) is the real advance over downloading a PDF programme, and it rewards specific prompts. Experience level, equipment, days available, injuries: the more of that in the first message, the less generic the plan.
Even the big logging apps noticed; Hevy pipes ChatGPT into its routine builder these days. The text-in, plan-out part of training has been absorbed.
Where the plan stops helping
It has never seen you move
The plan says "Romanian deadlift, 3×10". The model can describe a hip hinge in words, and words are the weakest possible way to learn one. It cannot see your rounded back, your knees caving at the bottom of a squat, the half-rep you count as a rep when tired. Every exercise name in that plan is a movement you still have to learn from somewhere, and for most people that somewhere is video: a coach on YouTube breaking down the hinge, a creator on Instagram showing the cue that finally clicks. The AI plan arrives with none of that attached.
The numbers are guesses
Ask for starting weights and you get hedged placeholders, sensibly so, since the model knows nothing about your body. Fine for week one. The problem compounds over weeks: progression in a programme depends on what you lifted last time, and ChatGPT has no record of that unless you paste your training history into the chat every session. People try exactly this, and the conversation turns into a scroll of numbers the model occasionally misremembers. By week four the "adaptive AI programme" is you, doing data entry into a chat box.
The plan is a wall of text
A chat transcript has no structure a gym session can use. Which day is today? What did the plan say about rest periods? You are scrolling a conversation, between sets, past your own prompts, looking for Tuesday. Copying it to Notes helps until the plan changes, and then two versions exist and you train from the wrong one. A programme needs a shape (days, slots, order) that a conversation never holds for long. The same failure sinks workout templates built in Notion, and a chat log is a worse home for a plan than a database.
Nothing closes the loop
Good programming is a cycle: plan, train, record, adjust. ChatGPT covers the first step well and the last step badly, and leaves the middle two entirely to you. It cannot know you hit 80kg for 8, that your volume fell this week, that a lift has stalled for a month. Without a log feeding it, the adjustment step runs on your memory, and memory is a poor training log.
Making the AI plan runnable
The fix follows from the failures, and none of it means abandoning the chatbot.
Interrogate properly. Give it your training age, equipment, schedule and injury history in the first prompt, then push back on the draft. Ask why it chose the volume it chose. Ask for a swap for any movement you lack equipment for. The redraft loop is where it earns its keep.
Attach a video to every lift. For each exercise in the final plan, find one demonstration from a coach you trust and keep it with the plan, so "Bulgarian split squat" is a thing you can watch at the rack rather than parse from a sentence. Building a programme out of videos covers how to pick the good ones.
Log outside the chat. Sets, reps and weights go into something built for recording them: a notebook does it, a logger app does it with PRs and weekly volume computed for you.
This is the gap Fitness Vault was built for, and it slots in without replacing the AI step. The programme ChatGPT wrote becomes a programme in the app: each slot holds the video you chose for that lift, imported from YouTube, Instagram, TikTok or Facebook, and while you train, the set logger sits on the same screen as the demonstration. The free tier (20 videos, one programme, manual tagging, full stats) fits a complete AI-written block. Where we are weak, for balance: if your plan is pure barbell work and you already know every lift cold, a plain logger like Strong or Hevy is the quicker tool, and Hevy's built-in GPT keeps the whole loop in one app.
The bottom line
ChatGPT compresses programme design from a £30 PDF to a thirty-second prompt, and for standard goals the output holds up. Treat it as the planning layer and judge it there. The training still happens in a world the chat window cannot see: movements you learn from video, numbers that need recording, a week that needs structure. Keep the AI for the thinking, give the plan a home with videos and a log, and the thirty-second programme turns into something you can actually run for twelve weeks.