Describe it in plain English. Get a real strategy file.

Why a prompt works better than a tutorial

Most strategies fail for one boring reason: the code is written against libraries the platform does not allow, or it manages its own positions instead of emitting signals. An AI has no way to know that unless it is told. This prompt states the contract up front — the imports that are rejected, the functions that are banned, the exact indicator library, and the rule that the strategy produces signals while the engine handles sizing, stops and profit.

What the prompt pins down

• The one rule that decides everything: your strategy returns two boolean series — when to go long, when to go short. The engine owns position sizing, stop-loss, take-profit, trailing and costs.
• The complete import allowlist and the denied-call list, so the model never reaches for os, sys, requests, sklearn, talib or getattr.
• The full indicator library with real signatures, including the ones that return tuples: macd returns three series, bbands and stochastic return two.
• The config keys the engine reads, and the instruction to never write into them — risk and exits stay yours, not the strategy’s.
• A required META block, so every tunable you declare shows up in the research sweep instead of being hard-coded.
• A self-check list the model runs before answering, plus a hard ban on inventing backtest numbers it never ran.