ChatGPT Trading Prompts: Three Crypto Templates and What GPT Does With Them

A ChatGPT trading prompt has two jobs: give the model numbers it did not have to invent, and force an output you can check. GPT does both well when told to, and drifts when not told. The three templates here are built from a prompt that runs in production every day, so the shape has been tested against a real model, on real prices, with the answers logged.

We run TradeRank.ai, a paper-trading league where language models manage separate portfolios in crypto and US equities. The OpenAI seat is GPT-5.6 Sol Pro. What follows is how it is prompted, what it wrote back over two weeks of decisions, and three prompts that carry the same structure into a ChatGPT window.

How the Arena Prompts GPT-5.6 Sol Pro

The call goes through OpenRouter to `openai/gpt-5.6-sol-pro` with `reasoning.effort` set to `medium`. The prompt is one long text block, sent together with the market data as a single user message: a mandate ("You are a medium-term investor, not a day trader"), the account state, hard constraints that mirror the validator, a description of the data, the model's own carried theses from the previous day, and a strict JSON output format. The market data follows it as JSON: a whole-universe feature table, candles for held and screened symbols, RSI per timeframe.

The server does not use a JSON-schema response mode. The prompt's rules demand "Output ONLY valid JSON — no markdown, no text before/after", and the parser pulls the first balanced `{...}` object out of whatever comes back. The prompt is not trusted either. Each decision goes through a validator that rejects anything under 0.80 confidence, any open without an `invalidation_price` on the losing side of entry, any add to a losing position, a second new position in the same cycle, and a reopen of a symbol closed that cycle. A rejected decision gets exactly one repair request, then the cycle finalizes. The full pipeline is on how it works.

What GPT Wrote Back: Fourteen Logged Cycles

The committed decision log for the OpenAI seat lags the live one; this post reads the first fourteen cycles of Season 8 in the committed log, 2026-08-15 to 2026-08-29 (the arena skipped 08-27 for every model). Counted from that file: 49 decisions, 8 opens, 2 closes, 39 holds, 0 rejections. Every open was a long. One open was crypto, BNB on the first day; the other seven were US equities. Nine of the fourteen cycles describe crypto as "overbought" somewhere in the reasoning. Those are counts for that window, not a verdict; the model page has the P&L.

Quotes are taken from the logged reasoning of each decision and the portfolio-level assessment, unedited.

It names a price for every thesis. Every hold cites a level: BNB "above 588" on 2026-08-16, BAC "unless BAC loses 61.8" on 2026-08-19, TMO with "556 remains the level that would negate the thesis" on 2026-08-23, MA with "569 now marking the breakout-support area" on 2026-08-25. The prompt requires an `invalidation_price` on every open and lets a hold move it; the model keeps quoting the number back.

It closes on the reason it gave for opening. The BNB close cites the target zone it had stated; the BAC close cites the breakout it had bought. The prompt carries each position's thesis forward every day under a PREVIOUS REVIEW heading, and GPT reads it.

It sits out crypto it calls overbought. From 2026-08-19 the reasoning calls crypto overbought in every cycle except 2026-08-28, several of those cycles say cash is preferable to a crypto entry, and every open after BNB was an equity. Nothing in the prompt says avoid high RSI; the prompt supplies `rsi_1d` and says "doing nothing is a valid decision". That is the model's own rule, and the reason the crypto templates below make the overbought question explicit rather than leaving it to the model's mood.

ChatGPT Crypto Trading Prompt 1: The Daily Review That Returns JSON

Use this once a day for a small crypto list. It copies the arena's contract: numbers in, one JSON object out, an invalidation price on every open, and hold as the default. Replace the bracketed data with values from your exchange or charting tool; do not ask the model to fetch them.

Paste the reply into a JSON viewer before you read it as prose. If a decision arrives without an `invalidation_price`, send it back with the one-line repair the arena uses: "Revise ONLY the rejected decisions or drop them. Reply with a full JSON object in this exact shape…" The repair never fired in the fourteen cycles above; every first answer fit the schema.

Template 2: Hold or Close, With the Level

This is the check the arena runs on every held position every day, lifted from the PREVIOUS REVIEW block. Give ChatGPT the thesis you wrote when you opened, the invalidation you wrote, and today's numbers, and make it say whether the thesis is still alive at a price.

The line that does the work is the rule against loosening the level toward entry. The arena's validator checks a level against the current price, not your entry: a long's level must sit below the current price, and nothing stops the model loosening a correctly sided level toward entry. In a chat window even that check is gone, so compare the returned number with the one you armed before you act on a hold.

Template 3: The Overbought Screen

GPT's own log shows it declining new crypto entries at daily RSI of 70 and above, and it does that unprompted. This template makes the same check the first step so a bullish read cannot skip it, and it asks for a table instead of JSON because the output is a list to scan.

GPT Habits to Plan For

It follows a schema better than it follows a rule. Zero rejections across the 49 decisions in the window is the schema working. The arena still validates every field, because a validator catches the day the schema does not hold. In the app, paste the schema into the prompt as above. In the API, add a strict JSON schema in `response_format` as well; the arena prompt predates that option and has not needed it.

Reasoning effort is a dial, and the arena sets it to medium. Higher effort costs time and money per call and did not get a trial in this data, so the log above is what medium looks like. If your plan exposes the setting, keep it fixed while you compare answers; changing it mid-test changes the model you are measuring.

Left to itself, it went long and sat in cash. Eight opens, all long, seven of them equities, across the fourteen logged cycles, nine of which call crypto overbought. The prompt allows shorts and says nothing about asset mix. If you want a short case considered, ask for it by name, as the bear-case template in the general prompts guide does. If you want crypto exposure, a screen like Template 3 will tell you when the model would rather wait, and you decide whether to overrule it.

Where to Go Next