The person I had in mind was 68, retired, comfortable with a browser and a spreadsheet, and completely uninterested in installing anything. If the answer to “how do I trade my signals automatically?” required opening a terminal, the answer was useless.
So I asked for something specific: a self-contained document containing everything an AI would need to interface with TradeDM and Alpaca on a client’s behalf and follow their trading instructions. One file. Hand it to an assistant; the assistant does the rest.
Grok wrote it. Then I showed that document to three other frontier AIs and asked each for its own version, adapted to its own platform.
None of them could produce one that worked.
The document
What came back is a one-page file Grok calls a Trade Card, in two parts.
Part 1 is a form the human fills in: your TradeDM key, your Alpaca paper key, which portfolio to trade, how much per ticker, and whether you will be back at 3:50 to close out. Part 2 is the recipe — the sequence the assistant follows, which the human never needs to read. It opens with a line I have grown fond of: “Human: ignore this part.”
You attach the file to a chat and say “execute today’s trade card.” That is the entire user experience. No install, no terminal, no Python, no scheduled job. Attach a file; read a summary of what filled.
And Grok can run it. It has a built-in execution environment that takes the attached file and makes the calls itself — check the keys, confirm the market is open, pull the signals, size the orders in whole shares, place them, print a plain-English fill list. The conversation is the app.
What the others said
I gave the same card to ChatGPT, Claude and Gemini and asked each to produce its own equivalent.
They understood it perfectly. Every one of them read the structure, understood what a signal was, worked out that an UP call means buy and a DOWN call means short, and knew to skip any ticker where two models disagreed. Gemini handed back a complete, well-organized version of the card with every step laid out properly.
It just could not run it. Neither could the others. Each one, in its own polite way, arrived at the same destination: here is a Python script — save it to your computer and run it.
That is a perfectly good answer for a developer. It is a wall for everyone else. You need Python installed. Your keys go in a file on disk. You have to run it at the right time of day. And when it fails at 9:31 you are reading a stack trace to find out why. My 68-year-old is not doing that, and should not have to.
What the test actually measured
Not intelligence. Every one of these models is easily smart enough for the job — it is a dozen HTTP calls and some rounding.
What it measured is whether the assistant can act on the user’s behalf inside the session. That is a product capability rather than a model capability, and it decided everything. The three that fell short were not reasoning worse. They were reasoning perfectly and then handing the work back.
Which leaves the part I keep chewing on. The brief was explicitly for a document that any AI could follow. The AI that wrote it is, so far, the only one that can.
Caveats, because they are real
September 2026. These products ship constantly and I would expect the gap to close, possibly before you read this. Nothing here is a verdict on anyone’s AI; it is a snapshot of which capability turned out to be load-bearing.
Paper trading. The card uses Alpaca paper keys, which trade play money.
Not us. TradeDM publishes signals. There is no order endpoint in our API and we never touch anyone’s account. What happens here is your assistant, your broker, your keys, your instruction — a distinction important enough that it gets its own post.
For the people who do want a script
We built one of those too: a small local day-bot with a real installer for macOS and Windows, for anyone who wants it running unattended with no chat window open. That is the other end of the same spectrum, and a fair amount more setup.
The point of the card was that most people should not need it.