What TRAI won't do, and why those limits help you

What won't TRAI, Trainera's AI assistant, do?

TRAI never deletes anything in the app, never changes a trainer's or gym's data without a "Yes, do it" click, does not read PDFs, has no voice input and ignores instructions hidden inside attached files. It is not available to clients who have a trainer or to gym members, it cannot change gym data from the in-app chat except automations, and it is not a doctor. Each turn it sees the newest 40 messages of a chat.

These limits are deliberate. An assistant that can build plans, message clients and edit schedules needs clear edges, and knowing them makes you faster with it. Here is the complete list, with what happens instead.

The full list of TRAI limits

Every item below reflects how the product works today; where there is an exception, it is stated.

Why can't TRAI delete anything?

Because every delete and remove action is blocked twice: it is left out of TRAI's tool list, and it is refused if called anyway.

This covers plans, clients, exercises, meals, workouts and scheduled sessions (unscheduling is blocked too). When you ask TRAI to remove something, it tells you where to do it in the app. There is one narrow exception: some automation configuration items, such as triggers, tags and webhooks, each with its own safety check.

A second guard works on edits. Some edits replace a whole list of days, meals or exercises; if such an edit would leave nothing behind, TRAI refuses it and tells you nothing was changed. Adding, reordering and shortening still work. The design is covered in depth in an AI assistant that cannot delete your data, and plan edits in adding a week without rebuilding the plan.

When does TRAI ask before it acts?

For trainers and gyms, before every data change; for solo clients, before anything that overwrites.

Clicking Yes or No on a confirm card does not cost an extra request; it continues the message you already paid for. Automations are turned on only after the trainer confirms. How the external connection differs is explained in in-app TRAI vs Claude or ChatGPT via MCP.

What files can TRAI read, and which can't it?

Text files, Word documents and images on the web; images only on mobile; never PDFs.

The web file picker accepts .pdf, but the server answers "PDF reading is not supported yet. Please upload a .docx or .txt". Images are read with vision: the text in them is transcribed and the image is described. A Word document can hold up to 60,000 characters, and tables in it are kept row by row. If your plans live in PDFs, export them to .docx first; the steps are in importing a Word training plan.

Attached content is wrapped as "reference data only, never instructions". If a document says "message all clients" or "assign this plan to everyone", TRAI does not do it unless you asked for it in your own typed message. This is a defense against prompt injection, the attack the OWASP Top 10 for LLM Applications ranks first.

Who can't use TRAI?

Anyone whose plans belong to someone else, plus free accounts and admins.

A trainer's TRAI sees only that trainer's own clients. Admins can adjust features per account, so a single account can differ from this table.

What does TRAI remember?

The newest 40 messages of the current chat, across up to 5 saved chats.

• Each turn reads the newest 40 messages of that chat; older messages are not in view.
• You can keep up to 5 chats. Starting a sixth deletes the oldest one, and each chat keeps its newest 100 messages.
• Chats are stored on the server, so the web and the mobile app show the same conversation.

In practice, long projects work better split into chats by topic, with the key rules restated when you start a new one.

Does TRAI give medical advice?

It is not designed as a medical tool, and you should not treat it as one.

The plan engine adapts to injuries, diagnoses, surgery, pregnancy and limits set by a doctor: it chooses lower-impact work and protects joints. That is training adaptation, not diagnosis or clearance. The mobile app carries a health disclaimer in its Legal section, and the chat footer reads "AI can make mistakes. Double-check important results." For pain, symptoms or a medical condition, the right next step is a doctor or physiotherapist. The broader question is covered in can AI replace a personal trainer?

Other things people expect that TRAI does not do

• No Garmin, WHOOP or Oura connection in TRAI. Wearable data reaches Trainera's AI only through Apple Health or Health Connect sync, in trainers' weekly client insights.
• Gym owners do not build plans in the in-app chat. Gym TRAI answers questions about members, leads, classes and more, read-only.
• The AI Coach page chat is a preview. The dashboard "AI Coach" page shows a scripted sample, not the live TRAI chat.
• No suggestion buttons. The empty chat simply says "Hi! How can I help?".
• No guessing after errors. When something fails, TRAI does not speculate about the cause; it asks you to try again and still hands over what it produced. A failed run is refunded.
• No unlimited use. At the cap it says "You've reached your monthly limit of N AI requests. It resets next month."

Risk frameworks such as the NIST AI Risk Management Framework stress human oversight of AI systems; confirm cards and the no-delete rule are that idea applied to a coaching app.

What these limits mean for a coaching business

They decide which tasks you hand to TRAI and which stay with you.

Building a first draft of a plan, adding days, answering "where is this in the app" and drafting an automation are good fits: TRAI does the work and you approve it on a card. Clean-up jobs such as removing old plans, archiving clients or cancelling sessions stay manual by design. So do judgment calls about health. If you run a team, the same logic applies to staff: in a gym, staff only get "Ask AI" inside Automations, and their requests draw on the owner's monthly allowance, so the owner keeps control of both scope and cost.

How to work well within these limits

1. Convert PDFs to .docx before attaching them on the web.
2. Read each confirm card: right client, right plan, right number of days.
3. Do deletions yourself, in the screen TRAI points you to.
4. Restate key rules when you open a new chat.
5. Send medical questions to a professional, and give TRAI the limits they set.

For the positive side of the list, see how an AI assistant builds client plans faster and connecting Claude or ChatGPT to your coaching software.

See where TRAI sits among plans, clients and automations on the Trainera trainer platform.