1. Describe a complete trading rule
An instruction such as ‘trade Bitcoin profitably’ leaves the important decisions undefined. Specify the market, evaluation interval, entry condition and exit condition. Also say whether the strategy should start in cash or establish a position immediately when a condition is already true.
Try this example prompt: ‘Use Bitcoin daily candles. Start in cash. At each daily close, enter long if the close crosses above its 100-day simple moving average. Exit if it crosses below. Use only completed candles and include trading fees in the backtest.’
This is an example of a testable rule, not a strategy we are claiming is profitable. Choose the historical date range and simulated capital explicitly when preparing the test.
2. Check what the AI built
Premiss generates real Python from the prompt. Ask it to explain the entry and exit conditions, then compare them with your original rule. ‘Crosses above’ and ‘is above’ are different conditions; a strategy starting in cash can behave differently under each.
Check that the moving average uses only information available at the time. Establish when a completed-candle decision becomes a simulated fill. Review indicator warm-up, missing data and position sizing rather than assuming the generated defaults match your intention.
3. Backtest the trades, costs and drawdowns
Run the strategy against historical candles and inspect individual entries and exits alongside the summary. A return percentage alone cannot tell you whether the rule is useful. A small number of lucky trades can dominate a result.
Include costs and compare the strategy with simply holding Bitcoin over the same period. Keep a separate period for evaluation after choosing your rules; repeatedly tuning the strategy to the same history makes the test less informative.
- Check the largest drawdown and how long recovery takes.
- Inspect trade count, turnover and sensitivity to fees.
- Review performance across rising, falling and sideways markets.
- Keep a record of the prompt, code version, date range and assumptions.
4. Run a paper bot on new market data
Start a hosted paper bot with the strategy you reviewed. It evaluates new market data and records simulated decisions and trades, so you can compare its behavior with the intended rules without placing real orders through that paper simulation.
For the daily moving-average example, a day without a crossover can correctly produce no signal. Check the evaluation history before interpreting silence as a broken bot. Paper trading still cannot reproduce every delay, rejected order or execution cost of a real account.
5. Choose where the bot sends its signals
On Pro and Quant, a paper bot can supply position-change signals to a supported destination. Choose Discord to post notifications to your own channel. Choose the 3Commas, OKX or TradersPost preset for a receiving bot or broker workflow, and follow that integration’s setup guide.
Configure sizing and execution in the receiving service. Premiss’s simulated position size is not an instruction to place the same order in your account. Check the instrument mapping and use a receiver’s demo or paper environment where supported.
The current signals cover entering and exiting long positions. Connecting a destination does not replay past trades or synchronize an existing position. Watch for the next position change, then check delivery history and the receiver separately: an accepted message is not proof of a filled order.