How NotchFit's AI Decides Your Next Workout

Most workout apps that call themselves "AI-powered" can generate a plan. Very few can explain why it contains what it does. You finish onboarding, hit a button, and a week of workouts appears — but there's no way to know whether the AI built something around your actual situation or just pattern-matched a plausible-looking template and dropped your name at the top. That ambiguity is a real problem. If you don't understand why your workout looks the way it does, you can't tell whether it's working, or push back intelligently when something feels off.

So here's what's actually happening inside NotchFit's generation — the inputs, the logic, and the parts that still don't work as well as I'd like.

What the AI starts with

When NotchFit builds your next week's plan, it starts from a structured picture of you assembled from onboarding and any history you've accumulated:

• Goals — strength, endurance, weight loss, general fitness, or a combination. These shape which exercises show up and what rep ranges the plan targets.
• Training days and session length — how many days you have available this week and how long each session can realistically be. A 30-minute session gets a different structure than a 60-minute one. The plan respects both constraints rather than assuming you'll find extra time somewhere.
• Location and equipment — home, gym, outdoor, or mixed. This determines which exercises are even on the table. If you train at home without a bench, incline dumbbell press won't appear. If you're at a hotel gym with only dumbbells, the plan stays in that lane.
• Fitness level and injuries — these gate exercise complexity and load ranges. A beginner and someone with two years of consistent training don't get the same plan even if everything else is identical.
• AI memories — if you've been using NotchFit for more than one cycle, your behavioral patterns and stated preferences travel into the generation prompt alongside your profile. The first week is an educated guess; by week eight the AI knows things about you that weren't in your original answers.

This is the foundation. Everything that follows builds on top of it.

The generation loop

NotchFit doesn't generate a plan in a single pass. The system runs a generate-review-revise loop, and the rough version of how that works is worth knowing.

First, the AI produces a draft plan from your full profile. Then a second pass — a reviewer — checks that draft against a set of rules: Are the exercise counts right for the session length and goals? Do the rep ranges match the intended training effect (strength work looks different from endurance)? Do dumbbell, kettlebell, cable, and machine exercises include specific weights rather than leaving them vague? Does the plan stay within your equipment constraints and location-appropriate weight caps?

If the draft doesn't pass, it goes back for revision with specific feedback about what needs to change. This can happen up to three times. The best-scoring version is what you see.

The practical effect: you're less likely to get a plan that's technically complete but obviously wrong for your setup. The reviewer catches errors before they reach you — the kind of thing that would have slipped through a single-pass system.

The quality of a plan is bounded by the quality of what the AI knows about you. Week one is an educated guess. Week ten is something closer to a real picture.

What it learns week to week

The generation loop starts fresh each week, but it doesn't start from zero. If you've been training with NotchFit for a while, your history travels with you.

The system tracks behavioral patterns across sessions — exercises you consistently skip, weight adjustments you make relative to what was prescribed, workout types you complete versus abandon. These observations build up over time and feed back into subsequent plans. If you've been using heavier weights than suggested on goblet squats for three weeks running, the AI eventually stops underestimating you on that movement.

You can also tell Sage, NotchFit's in-app AI, your preferences directly. "I don't like Romanian deadlifts." "I prefer upper/lower splits." "I have more time on Tuesdays." These go in immediately at full confidence and stay until you ask Sage to forget them. Your Workout App Should Remember You covers the full memory mechanics — how observations accumulate, how they fade when behavior changes, and how you can see everything the system currently knows.

For a deeper technical breakdown of the methodology — training principles, progressive overload tracking, how phase progression works — the methodology page has the full picture.

What it still gets wrong

A fair transparency disclosure means naming the limits, not just the capabilities.

The AI doesn't know how your body feels today. It can see that you logged a hard session on Thursday and that it's now Saturday, but it can't feel the residual fatigue in your legs or the shoulder tightness that's been building since Tuesday. If something hurts, you still need your own judgment. Sage can help you swap or modify exercises in the moment, but that's reactive — you still have to flag the problem.

The AI can't verify your form. It can give you a well-structured progressive overload plan, but it has no way to know whether you're performing a deadlift safely. Sound programming and sound execution are separate problems, and the AI only addresses the first one.

The AI's understanding of your equipment is exactly as accurate as what you told it. If you said "home gym" but you have a full barbell setup, the plan will underutilize what's available. If you said "dumbbells up to 50 lbs" but you're consistently lifting heavier, the suggested weights will lag until the memory system catches up to the pattern. The system learns, but it takes a few cycles to accumulate enough signal to adjust confidently.

And the AI is only as good as its exercise library. Occasionally it will suggest an exercise that's awkward for your body mechanics or simply not a good fit, for reasons that aren't captured in any of the data it has. That's what the skip-and-note flow is for — skip it, tell Sage why, and the system won't repeat the mistake.

What you can do with this

Understanding how the generation works changes how you interact with it.

The most efficient thing you can do in the first few weeks is give the system accurate information and correct it quickly when it's wrong. If a weight feels off, adjust it in the app so the pattern gets recorded. If an exercise is a consistent non-starter for you, tell Sage rather than silently skipping it — that converts an implicit signal into an explicit memory faster.

If a plan looks wrong in a way you can't easily fix by editing, it's usually faster to tell Sage what specifically isn't working than to modify workouts one by one. "I don't have access to cables this week" or "these sets feel too light for my level on this movement" gives the AI something to work with. A targeted conversation with Sage will get you closer than manual tweaks alone.

And if you want to know what the system currently has on file about you, the AI Memory section on your Profile page shows every observation and stated preference, labeled by source — what you said explicitly, what Sage inferred from conversation, and what was observed from your workout behavior. Nothing is hidden.

The system is designed to get more accurate over time, not to be perfect from day one. That means the first plan is a starting point, and every subsequent plan is an improvement — as long as you give the system something to learn from.