# AI workout plans built from real exercise videos > TRAI does not write exercises freehand. The AI writes a prescription and a deterministic engine assembles the plan from more than 1,800 library exercises with images and videos, picking the trainer's own exercises first. *Author: Trainera Team | Published: 2026-10-11 | Reading time: 8 min* ## Why is an AI plan built from an exercise library better than one a chatbot writes? Because every exercise in it is a real library entry with an image and a demo video, not a name the AI made up. In Trainera, TRAI does not write workouts freehand: the AI writes a structured prescription (goal, level, equipment, days, injuries, instructions) and a deterministic engine assembles the plan from a library of more than 1,800 exercises, picking the trainer's own exercises first. The client opens the plan and can watch how each movement is done. A general chatbot can write a good-looking plan as text. What it usually cannot do is link each line to a demo video, respect a fixed equipment list across every session and save the result where the client trains. That gap is what this post is about. ## How does the prescription-plus-library approach work? The work is split into two steps, and the AI only does the first one. | Step | Done by | Output | | -------------------------- | ---------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | 1\. Understand the request | The AI model | A structured prescription: goal, experience level, equipment, number of workouts, injuries and free-text instructions (split, day kinds, session order, implements, session length, wanted exercises) | | 2\. Assemble the plan | Trainera's deterministic plan engine | Days built from the exercise library, each session ordered in blocks: optional cardio at the start, warm-up, main work, optional cardio at the end, stretching | | 3\. Save | The server, after your confirmation (trainers) | A plan in the library or on the client's account, with images and videos attached | The point of the split: the AI does what it is good at, turning "my knee hurts on lunges and I only have bands" into structured inputs, and the engine does what software is good at, picking only exercises that exist and fit those inputs. In the engine's own design notes, the AI does not invent exercises. This is the same engine as the Generate buttons in the app, and the same one TRAI uses for trainers and for solo clients on Client Pro and Premium. A deeper look at the steps is in [how AI builds your workout and diet plan](/blogs/how-ai-builds-your-workout-and-diet-plan). ## Chatbot-written plan vs library-built plan The two can read similarly on screen; the difference shows when the client trains with them. | Point of comparison | Plan written by a general chatbot | Plan built by TRAI and the engine | | ---------------------- | ----------------------------------------------------- | -------------------------------------------------------------------------- | | Exercise names | Free text; can include unclear or invented variations | Library entries with fixed names | | Demo media | None, or links the model suggests | Official image and video attached from the library | | Session order | Whatever the text says | Fixed blocks: cardio start, warm-up, main, cardio end, stretch | | Equipment | Depends on the model following the prompt | Equipment setting plus implements to use or avoid, applied by the engine | | Injuries and pregnancy | Depends on the prompt | Handled by the engine, which chooses lower-impact work and protects joints | | Where it lives | A chat window or a copied document | Saved in the app, ready to assign or train | | Your own exercises | Not known to the chatbot | Picked first from your library | Unreliable output from language models is a documented risk, not a hypothetical one. The [OWASP Top 10 for LLM Applications](https://owasp.org/www-project-top-10-for-large-language-model-applications/) lists misinformation and overreliance on model output among its risks, and the [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) covers how to manage such risks. Constraining the AI to a known library is one practical answer to that problem for workout plans. Our comparison with prompt examples is in [ChatGPT workout plans vs a real AI fitness app](/blogs/chatgpt-workout-plan-vs-ai-fitness-app). ## Why does the trainer's own library come first? Because a trainer's exercises carry their own cues, their own videos and the names their clients already know. When TRAI builds a plan for a trainer, the engine picks the trainer's own exercises and meals first. Only what is missing comes from Trainera's reference library. When a trainer dictates exercises "from my library", TRAI first reads the library overview and copies the names exactly; plan items are linked to library entries by name (case does not matter), which attaches the trainer's own images, videos and data. Everything a saved plan uses is added to the trainer's own library automatically, with images, and duplicates are prevented by name. The first time in a chat, TRAI's confirm card says that missing exercises and meals will come from the reference library and be added. Over time, then, the library grows with the plans you actually use. Example prompts for trainers (examples, adjust to your setup): * Example: "Build a 4-day hypertrophy plan for Client A from my own exercises where possible. Full gym, 60 minutes." * Example: "Make a 3-day home plan with my 'Band pull-apart' and 'Tempo goblet squat' in every session, the rest from the library." * Example: "Create one 30-minute workout: start with 5 minutes rowing, then legs, finish with stretching." ## One request, followed from message to saved plan Here is how a single example request moves through the system; the request is invented for illustration. A trainer writes: "3-day plan for Client B, home, dumbbells and a band only, 40 minutes, no jumping because of her ankle, finish every session with stretching." TRAI reads the client's profile and turns the message into a prescription along these lines: goal and level from the profile, a home equipment setting, the dumbbell-and-band limit and the stretching finish in the instructions, the ankle in injuries, three workout days. The confirm card states what will be created, how many days and for which client. After "Yes, do it", the engine selects exercises that fit those inputs, orders each session from warm-up through main work to stretching, attaches images and videos, and saves the plan to the trainer's library. That takes up to half a minute. The trainer then reviews it, and asks TRAI to assign it when it looks right. Nothing in that chain asked the AI to type an exercise name from memory. ## What exactly is in the library? More than 1,800 exercises and more than 1,100 meals, each with official images, and videos for the exercises. Different pages in the app describe the exercise count slightly differently as the library grows, and is above 1,800 today. The meals matter too: nutrition plans are built by the same engine from real library meals, with calories and macros. | Library part | Size | What is attached | Who it is for | | --------------------------------- | ---------------------------------------- | ---------------------------------------------------------------- | ---------------------------------- | | Reference exercises | More than 1,800 | Official image and demo video | Every plan built by the engine | | Reference meals | More than 1,100 | Official image; used with calories and macros in nutrition plans | Nutrition plans | | Trainer's own exercises and meals | Whatever the trainer creates or collects | The trainer's own media and data | That trainer's plans, picked first | The equipment and implement words the engine understands (machine, barbell, EZ-bar, dumbbell, kettlebell, band, bodyweight, pull-up or dip bar), along with warm-up, stretching, cardio and body-part terms, are covered in all 26 app languages. See [how to tell an AI coach what equipment you have](/blogs/trai-equipment-home-gym-ai-plan) and [day kinds and session order](/blogs/trai-day-kinds-session-order). ## When does TRAI write exercises itself? Only when you give it the exact content, or for small edits. * **Dictated plans:** when you list every exercise yourself, TRAI writes them in as given. * **Small edits:** swapping, adding or removing an exercise, or changing sets and reps on an existing day. * **Document import:** when a trainer attaches a .docx or text file with a plan, TRAI transcribes it as written. Exercises whose names match the reference library get the official image and video attached automatically. The import is covered in [importing a Word plan](/blogs/trai-import-word-training-plan). In all three cases the library link still helps: matched names get media, and saved items land in the trainer's library without duplicates. ## What does a library-built plan not solve? It fixes the exercise problem, not the coaching problem. * A demo video shows the movement; it does not watch your client do it. Form feedback is still the trainer's job. * The engine builds what the prescription says. A vague request gives a generic plan, so specific instructions matter. * Injury handling means lower-impact choices and joint protection, not medical clearance. The chat footer says "AI can make mistakes. Double-check important results." * TRAI never deletes anything and, for trainers, saves only after a confirm card. ## Who gets this and how much it costs Trainers on Starter and up, and solo clients on Client Pro and Premium. Each chat message is one request, and a plan built inside TRAI does not spend the Generate-button quota. | Plan (US prices) | Price per month | TRAI requests per month | | --------------------- | --------------- | ----------------------- | | Trainer Starter | $19.99 | 20 | | Trainer Pro | $49.99 | 50 | | Trainer Business | $99 | 100 | | Client Pro (solo) | $5.99 | 40 | | Client Premium (solo) | $6.99 | 90 | Prices differ by country on [trainera.fit/pricing](/pricing). Trainer Enterprise has 200 requests a month at a custom price. Free plans have no TRAI chat but include one AI training plan and one AI nutrition plan from the Generate buttons. For alternatives, see [the best AI workout generator apps](/blogs/best-ai-workout-generator-apps). ## How to get a better plan out of the library 1. Name the equipment exactly, and the implements you want or want to avoid. 2. State injuries in plain words; the engine reads them as written. 3. Trainers: build up your own library with your cues and videos, because it is picked first. 4. Say the session order if it matters ("cardio first, stretch at the end"). 5. Open one exercise video in the saved plan before assigning, to check it matches what you meant. _See what library-built AI plans look like on the [AI fitness coach page](/ai-fitness-coach)._ ## FAQ ### Does AI make up exercises in workout plans? General chatbots can. In Trainera, the AI writes a structured prescription and a deterministic engine picks the exercises from a library, so the AI does not invent exercises in engine builds. ### How many exercises are in Trainera's library? More than 1,800 exercises with official images and demo videos, plus more than 1,100 meals. ### Can the AI use my own exercises and videos? Yes. For trainers the engine picks your own exercises and meals first, and items matched to your library carry your own images and videos. ### Is a ChatGPT workout plan as good as an app plan? The text can be similar, but a chatbot plan has no linked demo videos, no fixed equipment rules and is not saved where the client trains. ### Do exercises from AI plans get added to my library? Yes. Exercises and meals in a saved plan are added to the trainer's own library automatically, with images and without duplicates. --- Source: https://trainera.fit/blogs/trai-library-exercises-real-videos