Skip to main content

AI assistant

StrongFish can put your training in front of a language model: a chat that answers as a powerlifting coach, a reading of a session you have just finished, and a reading of the last ninety days of your fitness data.

It does not come with a model. You connect your own account at a provider - the key is yours, the bill is yours - and the app spends it on your behalf. You can also lend your connection to other members, the way you lend a video storage.

Connecting a model​

In Settings, under AI connections, add one:

FieldWhat it is
NameWhat you call it. Left blank, it is named after the model - claude-sonnet-5-5-a9bffea3 - which is what the field offers while you type.
ProviderOpenAI, Mistral, Anthropic (Claude), Google Gemini, OpenRouter or DeepSeek. This is the wire format, not a commitment to the vendor.
AddressWhere the requests go. It is filled in for the provider you pick, and it is editable - point it at a LiteLLM, an Ollama, an OpenRouter account, anything that speaks the same protocol.
ModelWhich model to ask. The field suggests the ones your provider is known to serve, and takes any name you type - a model released yesterday works the day it is released.
API keyThe key from your provider. It is kept on the server and never sent back to a browser or a phone - once saved, the field offers to replace it rather than showing it.

Every placeholder follows the provider you pick, so what the fields show is that provider's own address, model and key format.

One connection is enough. If you have several - your own and one your coach lent you - you are asked which one to spend: a picker above the chat, and a menu on each AI button. The app never guesses when there is a choice, and never asks when there is not.

Which one answers by default​

Your connections are a list, and the first one answers when you do not pick - it is also the first row of the menu on every AI button. The arrows beside each one move it; a connection somebody shared with you sits at the end and stays where its owner put it.

A connection you have just added goes to the bottom, so pasting a key never quietly changes which subscription your questions are spent on.

Lending it to other members​

The group icon beside a connection opens its access list. Anybody you add there can ask questions through it, and nothing else: they never see the key, they cannot edit the connection, and they cannot share it on. Revoking takes effect immediately.

This is how a coach pays for one key and lets a roster of athletes use it.

The coach chat​

AI coach in the menu. Ask anything about your lifting - programming, technique, recovery, meet day. The model answers in the language your account is set to.

It already knows who you are. Every question carries your file: your age, bodyweight and height, every 1RM you have recorded, the block you are running with everything you have logged against it, and what your watch counted in the last twenty-four hours. So "should I squat tonight?" is answered against tonight's session and today's step count, not in general - and the weights it talks about are the kilos your program actually prescribes.

The file is rebuilt for each question rather than written into the conversation, so an old thread never answers from a max you have since moved.

Which AI answers​

If you have more than one subscription, the send button asks which one - it opens them by name, the same way the AI buttons on a session do. With one subscription it simply sends.

The choice is per message, not per conversation: one thread can run on your own key for one question and your coach's for the next. Each subscription points at its own model, so picking a name picks the model with it.

That is why every answer says what produced it and when, in small italics underneath: Generated by claude-opus-5-5 on November 2nd of 2026. The exported conversation carries the same line under each answer.

Every exchange is kept, so the list on the left is your history rather than this tab's. For each conversation you can:

  • rename it - the title starts as your first question;
  • share a link - anybody holding it can read the thread, the way a shared ChatGPT conversation works. Press the link button again to revoke it, and the link stops working at once;
  • export it as a PDF or a picture, to keep or to send on;
  • delete it - one at a time, several at once with the checkboxes, or the whole history.

A reading of one session​

Open an assigned program, open a session, and the AI feedback button sits under its sets.

It is off until the session is finished - every set logged. A reading of three sets out of eight is a reading of a session that did not happen that way, and it would be taken for a verdict on your training rather than on an unfinished record of it.

What the model is shown is the session sheet itself, the same document you can export: the prescription on the left, and what you actually did on the right - reps, load, RPE felt, estimated 1RM, and your own comment on each set. The sessions already completed earlier in the block go with it, so the reading can say whether you are progressing, holding or drifting rather than describing one day in isolation.

The answer is kept: it costs money to produce, asking twice gives two different readings of the same session, and you should be able to come back to it in March.

You can ask as many times as you like, and the readings stack up rather than replacing one another - a second opinion from a bigger model, or your coach's reading beside your own. Each one says underneath it which subscription produced it and when ("Generated by my Mistral key on October 19th of 2026"), which is how you tell two apart when they disagree.

Deleting one asks you to confirm. It cost money and it does not come back - you would have to ask for a new one.

Your coach sees it​

If the program belongs to a club, the reading is part of the session's record - your coach reads it the way they read your comments.

A coach can also generate it for you, from their own AI connection. That is what lets one subscription serve athletes who have none.

In the exports​

A session that has a reading carries it into the PDF, the spreadsheet and the picture, printed under the table it is about.

A reading of your fitness data​

On the Fitness data screen, under the charts, the same button reads what your watch and scale recorded: up to ninety days of steps and heart rate, and your latest body composition.

Ninety days because the question is a direction rather than a day - which weeks were busy, which were quiet, whether the resting figures are drifting.

Ask for it whenever you like, and - like the AI buttons on a session - the button is where you pick which subscription reads it, when you have more than one.

Each reading replaces the one before it, rather than stacking up as a session's do: what it describes is a window that keeps moving, so an older reading of it is out of date rather than a second opinion. The line underneath says which subscription produced the one you are looking at, and when.

It goes into the exported health sheet too, PDF and picture alike, under the charts it is about - as long as it is less than a day old. An older reading stays on screen with its date, but is left out of the document, where it would read as a comment on numbers it never saw.

What the model is told​

The instructions are markdown files on the server, not strings buried in the code, so they can be improved without touching the app:

FileWhat it governs
coach.skill.mdHow a coach answers at all: the voice, what to do with an RPE, what never to claim.
feedback.skill.mdWhat a session review specifically is.
health.skill.mdWhat a reading of fitness data specifically is - including that it is not a diagnosis.

The providers themselves are a table in the database rather than a list in the code, so an operator can correct an address, add an OpenAI-compatible service or teach the form a new model name without a new release.

Each is sent with the question, the common one first. The answer is always asked for in the language your account is set to, English when nothing is.

What it is not​

The model is not a doctor and not your coach. It reads what is in front of it - a sheet of sets, a series of step counts - and it says what a well-read coach would say about them. It does not know about the knee you did not log, and it should never be the only reader of your training.