Prompt Engineering Review
A trading agent's prompt is an operating constitution. It determines what evidence matters, how uncertainty is represented, when tools are called, and how conviction becomes exposure.
The strongest prompts specify decision processes rather than predicted answers. These ten patterns can be evaluated through completed BotTrade runs that expose their decisions, trades, returns, and risk.
Require the agent to classify the observed market structure before evaluating individual symbols. The response should name the evidence, confidence, expected duration, and conditions that would change the classification. In BotTrade, the prompt can use visible bars and scanner output, then produce a bounded portfolio decision whose later trades reveal whether the regime label was useful.
Separate the response into observed facts, calculated features, interpretations, assumptions, and invalidation conditions. This prevents a model-generated narrative from being restated as market evidence. The final BotTrade decision should cite the relevant ledger entries and omit unsupported claims. After completion, the trade record shows whether the thesis survived subsequent observations or required revision.
Frame every proposed order as a change to total exposure, cash, concentration, and existing positions. The agent must explain what the portfolio owns before and after the action, not merely why one symbol looks attractive. BotTrade provides explicit cash and position state, allowing the prompt to produce orders that respect the scenario's leverage, universe, and shorting constraints.
Ask a critic to construct the strongest case against the proposed action, identify missing evidence, and test the invalidation logic. Limit the review to one pass so debate does not consume the decision window. The trader must either revise the order or explain why the objection does not change it. Store both views with the BotTrade run for later attribution.
Translate conviction into a maximum capital allocation through explicit rules for concentration, leverage, loss tolerance, and available cash. Confidence should never map directly to an unlimited position. Require the model to calculate post-trade exposure before submitting an order. BotTrade's portfolio state and scenario constraints make the resulting size decision visible and reject actions that violate hard rules.
Define when the agent should scan the universe, inspect a limited symbol set, use external research, or act with current evidence. Each tool call should answer a named question and have a stopping condition. BotTrade's scan_market and inspect_symbols tools support this staged workflow, reducing context use and making it easier to identify redundant research loops.
Provide a compact record of earlier theses, submitted actions, and unresolved invalidation conditions rather than a transcript of every turn. Ask the agent to state what changed since its previous decision. BotTrade supplies the current portfolio and simulated time, while the application retains the decision memory. This structure helps expose repeated mistakes without leaking later market information.
Organize analysis around a catalyst, the instruments it may affect, expected transmission channels, timing, alternative outcomes, and evidence that the event is already priced. The final response must convert that map into a position or an explicit hold. Historical BotTrade scenarios let builders inspect whether the agent reacted before, during, or after the relevant market repricing.
Allow broad research internally, then require one compact output containing the current thesis, portfolio impact, orders, and hold or exit condition. Compression reduces contradictory instructions and makes tool arguments easier to validate. BotTrade's typed decision helpers and MCP order schema provide concrete output targets, so verbosity cannot conceal an absent or malformed action.
After a run or closed trade, ask which observations supported the thesis, which assumption failed, how sizing affected the result, and what rule should change. Do not let the model rewrite its earlier rationale. Use the stored BotTrade decisions and get_trades evidence as the source record, then preserve the attribution as input to the next engineering revision.
BotTrade allows builders to change a prompt pattern, run the agent through the market simulator, and inspect the resulting return, risk, trades, and portfolio behavior.