How TrainForge uses AI in your training

2026-05-23 · 7 min read

Not the hype version. A specific walkthrough of where the AI shows up in your program, what it does well, what it doesn't try to do, and what's coming next.

"AI-powered" appears on every fitness app's marketing page and means almost nothing. So instead of the hype version, here's exactly where AI shows up in TrainForge, how it works under the hood, where it adds real value, and the places we deliberately don't use it.

1. Program generation that respects your equipment

When you finish onboarding, the AI builds you a personalized program in about 30 seconds. The reason it works is that it doesn't pick exercises blindly — it picks from your actual available equipment. A Planet Fitness member gets fixed barbells 10–60 lbs, dumbbells 5–75, and specific machines. A home-gym member gets whatever's in their setup. A bodyweight-only user gets a program where every exercise needs nothing but the floor.

Most "AI program generators" pick from a generic exercise library and hope you can adapt. TrainForge feeds the AI your exercise list directly, so the program it returns is guaranteed executable.

Beyond equipment, the AI considers your goal (fat loss, muscle, recomp, strength, etc.), preferred split (PPL, Upper/Lower, Full Body, Bro Split), days per week, time per session, focus areas, and any injuries. It places compound lifts first, balances opposing muscle groups, programs in deload weeks every 4 weeks, and writes a brief rationale explaining why the program looks the way it does.

2. Food logging that doesn't require typing

Manual food logging is the single biggest reason people stop tracking nutrition. TrainForge offers three lower-friction options, all AI-powered:

Text-to-macros. Type "two scrambled eggs, half an avocado, slice of toast." The AI returns calories, protein, carbs, fat, and fiber in about two seconds.
Photo-to-macros. Take a picture of your meal. The AI identifies the foods, estimates portion sizes from visual cues, and returns the same macro breakdown.
Barcode scan. Not AI — uses Open Food Facts for the lookup — but worth mentioning as the third option.

None of these will match a digital food scale for precision. They're optimized for consistency over accuracy: an estimate you actually log every day produces better outcomes than a precise measurement you skip three times a week.

3. Weekly coaching that watches your data

Every week, the AI reviews your training data: did you hit your sessions, are your lifts progressing, are you hitting calorie and protein targets, is your weight trending the direction your goal suggests? If something's drifting, you get specific recommendations:

"Bench press hasn't moved in 3 weeks at 3x8 @ 165. Try double-progression — drop to 3x6 @ 175 and add a rep each week."
"You've been at a 600-cal deficit for 4 weeks. Considering a 1-week diet break at maintenance to reset hormones."
"Squat volume + leg press volume is high. Consider swapping leg press for walking lunges this week to manage knee stress."

You can accept any suggestion with one tap and the AI rewrites the relevant part of your program. You can ignore it and nothing changes. The intent is to surface what a coach would notice if they reviewed your training journal weekly — not to push you around.

4. Adaptive macros that learn your real metabolism

Every calorie calculator on the internet is wrong for you specifically, because formulas predict average people and you're not average. TrainForge starts with a formula estimate, then switches to a measured TDEE once you've logged food for 5+ days and weighed in at least twice. The system computes your actual maintenance calories from food intake minus weight change, then re-targets your daily calories to match your goal.

It runs weekly. If your weight trend is moving faster than your goal expects, calories adjust up. If it's stalled, calories adjust down. The math is detailed in the calorie-math article.

5. Trainer summaries (gym version)

Trainers managing 30 members can't read every workout log. TrainForge generates a one-paragraph AI summary per member that highlights: progression rate, session adherence, recent PRs, training issues, and nutrition adherence. The trainer sees the summary, decides whether to dig deeper, and saves hours per week of manual review.

Where we deliberately don't use AI

Some places AI looks like it would help but actually hurts. TrainForge avoids:

Form coaching from video. The tech exists, but the false-positive rate is high enough that bad advice ("your knees are caving") can cause injury. We'd rather not give cues at all than give wrong ones. This may change as multimodal models improve.
Conversational chatbot coaching. A coach you can text 24/7 sounds great until it tells someone with sciatica to "just deadlift through it." Coaching advice goes through structured prompts with safety guardrails, not a freeform chat where the model can wander.
"AI personal trainer" replacing humans. A trainer who knows your shoulder injury, watches your form, and notices the weeks you're not yourself is irreplaceable. AI handles the data and patterns; humans handle the relationship.
Hidden algorithmic decisions. Every AI recommendation is shown with reasoning, and you can override or ignore it. The model is a coach, not a black box dictating your program.

The model under the hood

For the curious: TrainForge uses OpenAI's gpt-4o-mini for all generative AI, with Structured Outputs to guarantee the response matches our schema (no malformed programs, no hallucinated exercise IDs). System prompts are versioned and tested against a persona rubric — we run quality evals against fake members (beginner, intermediate, advanced, injured) every time we change a prompt, and any score below 80% blocks the change.

Cost is tracked per gym and per user. The average member generates <$0.05/month in AI costs across program generation, food logging, weekly check-ins, and coaching combined. That's why we can bundle AI into every tier instead of charging separately for it.

What's coming next

Three things on the roadmap that we think will materially change how training works:

Wearable integration. Pulling sleep, heart rate variability, and resting heart rate from Apple Watch / Garmin / Whoop. When the AI sees you slept 5 hours and HRV is tanked, the day's program adjusts — drop the working weight 10%, kill the accessory volume, push the PR attempt to a better day.
Voice-logged sessions. Speak your sets between reps instead of tapping a phone screen with sweaty hands. The AI transcribes, parses, and logs in real time.
Long-context program memory. The next generation of models can hold your entire 2-year training history in context when making decisions. The AI will recommend programs based on what's actually worked for you historically, not generic patterns.

The honest bottom line

AI in fitness is at the "useful tool" stage, not the "replacement coach" stage. It can save you hours per week on the data work (food logging, program design, weekly review) and help you notice patterns you'd miss. It can't watch your bar path, talk you through a hard week, or know when life is the reason for the missed session. The version of training that combines both is meaningfully better than either alone — which is exactly the bet TrainForge is built on.