Most AI trading prompts ask for a prediction, and a prediction is the one thing a language model can't give you and will write anyway. Ask "will this go up?" and you get confident fiction. Ask for the price that proves the idea wrong, the risk-reward at that stop, and the conditions for no trade at all, and you get an answer you can hold up against your chart and check.
That's the whole shift. You stop asking the model what will happen and start asking it to commit to things that are either on the chart or aren't.
Why prediction prompts get fiction
A chatbot produces the most plausible continuation of what you typed. "Will NVDA break out today?" has a very plausible continuation: a balanced-sounding paragraph that leans one way, mentions momentum and a "key level," and closes with a caution. It reads like analysis. It would read the same if the stock were about to drop.
The problem isn't that the answer is wrong. It's that you have no way to find out before the trade does it for you. There's nothing in "likely to continue higher if momentum holds" that you can point at on a chart and say that's there or that isn't.
A constrained prompt changes what comes back. Not because the model got smarter, but because you asked for outputs that have a location.
Same chart, four ways to ask
| What you type | What comes back | Can you check it? |
|---|---|---|
| "Will this go up?" | A lean in one direction with reasons that fit either outcome | No. Only the trade can tell you |
| "Is this a good entry? How confident are you?" | A yes with a percentage attached | No. The percentage wasn't measured from anything |
| "Price is 42.00, support at 41.60. Where is a long wrong?" | A stop level and the reason for it | Yes. Find the level on your chart |
| "...and if risk-reward is under 2:1, say no trade" | A plan with the math shown, or a refusal | Yes. Redo the subtraction |
The level that proves you wrong
Invalidation is the price at which the reason for the trade no longer exists. If you're long because buyers defended a level, the idea is wrong when price closes through that level. It's not a prediction. It's a condition, and it's the most useful thing you can make a model state, for three reasons.
First, it's checkable. The model has to name a price and tie it to something visible: a swing low, the bottom of a range, the far side of a gap. You can hover that candle. If the level is floating 30 cents from any wick, you've caught a bad read before it cost anything.
Second, everything else hangs off it. Stop distance sets risk-reward, and risk-reward plus your account limit sets size. A plan without an invalidation isn't a plan with one field missing. It's a direction and a hope.
Third, it's the question that exposes a weak idea. If neither you nor the model can say where the trade is wrong, that tells you something about the trade.
The prices in that figure are a worked example, and the next section uses them.
Make it show the math
A model that says "this offers a favorable risk-reward" has told you nothing. A model that says "risk is 42.00 minus 41.50, reward is 43.00 minus 42.00" has given you two subtractions you can redo in your head. Ask for the second kind, always, and ask for the arithmetic rather than the conclusion. Language models make arithmetic slips, and a shown calculation is one you can catch.
Give it your account risk limit too, and it can carry the same numbers through to size.
The arithmetic the reply should contain
Notice what the 33% is and isn't. It's the win rate this setup needs just to break even, derived from the plan. It is not the chance the trade works. No prompt gets you that number, and a reply that offers one has made it up.
Give it permission to say no trade
Here's a quirk worth knowing. If your prompt says "give me a trade plan for this chart," a plan is the only shape the answer can take. The model will build one out of a chart that doesn't deserve one, because you asked. Messy, mid-range price action gets the same tidy entry, stop and target as a clean setup.
The fix is to make "no trade" a legal answer and to define when it applies. A minimum risk-reward is the simplest condition: if the target is less than twice the distance to the stop, there's no trade. Add the ones that match your own rules, like price sitting in the middle of a range, or the nearest invalidation being wider than your risk limit allows at a sensible size.
You'll know it's working when the model starts turning charts down. That's not the tool being unhelpful. Most charts, most of the time, aren't a trade.
The prompt, assembled
Put together, the structure looks like this. The bracketed parts are yours to fill in from your platform, not the model's to guess.
I'm looking at [ticker] on the [timeframe] chart. Current price is [price]. The levels I see are [support] and [resistance], and today's range is [low] to [high]. I risk at most [x]% of a [size] account per trade.
Using only these numbers, tell me: (1) the price at which a [long/short] idea is wrong, and what on the chart that level marks; (2) the entry and target, with risk and reward as subtractions and the resulting ratio; (3) the share size at my risk limit. If risk-reward is below 2:1 or price is mid-range, answer "no trade" and say why. List separately anything you're assuming that I didn't give you. Don't forecast direction or give a confidence figure.
Each sentence is there to close a specific hole.
Five constraints every trading prompt needs
-
Your numbers, typed in — Price, timeframe, levels and range from your own platform. The model has no feed, so anything it supplies is recalled or estimated.
-
Invalidation before entry — Ask where the idea is wrong and what that level marks on the chart. A stop with no reason attached is just a distance.
-
Risk-reward as shown arithmetic — Two subtractions and a ratio, written out. You redo them in five seconds, and you catch the slip if there is one.
-
A no-trade condition — A minimum ratio and the explicit instruction that declining is a complete answer. Otherwise every chart gets a plan.
-
Assumptions listed separately — Anything the model filled in that you didn't provide goes in its own list, so you can see exactly what to verify.
-
Asking how confident it is — The percentage comes from nowhere. It makes a weak plan feel measured, which is worse than no number at all.
If you'd rather attach a screenshot than type levels, keep the typed price and timeframe anyway and add "where the image and my numbers disagree, use mine." A picture gives the model the shape. It doesn't give it measurements, which is the gap we walked through in ChatGPT for day trading.
What a prompt can't fix
A good prompt changes the form of the answer. It doesn't change what the model knows. It still has no live quote, it can still misread a level, and the plan it returns describes the numbers you gave it at the moment you gave them. If price has moved by the time you're at the order ticket, the ratio in the chat is no longer the ratio in your account, and you recompute.
It also can't make the trade work. A well-formed plan with a clear invalidation loses regularly. That's what the stop is for. The structure doesn't promise a profit; it guarantees you know the size of the loss before you take it.
And you have to type it every time. That's the practical weakness of prompt-craft: the constraint you forget on a rushed morning is the one you needed. Purpose-built tools exist to hold the structure for you. An AI trading assistant or an AI trade plan generator returns entry, exit and risk fields on every chart, so the stop isn't something you have to remember to ask about. The checking is still yours. Any read of a chart, from any tool, is a draft until you've found its levels on your own screen.
The best prompt you'll write this week probably isn't clever. It's the one that hands over your own numbers, asks where you're wrong, and leaves room for the answer to be "not this one."