Running language-model agents in production for market analysis teaches a blunt lesson fast: the model is rarely the bottleneck. The prompt is. Here are the lessons that survive contact with real markets.
Feed data as narrative, not a dump
The instinct is to hand the model a table of numbers and let it sort them out. In practice, models reason better over framed data. “RSI is 72 — elevated, but consistent with the strong ADX of 31 and +DI leading -DI” gives the model the relationships, not just the values. A raw table makes the model re-derive context every time; a narrative carries it. AlphaFlowSeven assembles market data into a narrative section of each agent’s prompt for exactly this reason.
Enforce a response format
Free-text answers are unusable downstream. If you need a decision your system can act on, you must constrain the output — a strict schema like {"decision": "...", "confidence": 0.0, "rationale": "..."}. Two benefits: the output is parseable, and the act of requiring a confidence number forces the model to commit to a stance rather than hedging in prose. A response-format contract is as much a reasoning tool as a plumbing one.
Give each role a sharp mandate
Vague roles produce vague analysis. “Analyze the market” is worthless; “You are a Risk Analyst. Your only job is capital preservation. Identify what could go wrong with this position and quantify the downside” produces something useful. In a multi-agent council, sharp per-role mandates are what make the agents different — a Bull Advocate told to build the strongest long case will find arguments a neutral analyst glosses over.
Small changes, big effects
Prompts are high-leverage and brittle. A one-word change to a Lead Trader’s instruction can shift the tone of every future decision. This cuts both ways: it means prompts are powerful, and it means prompt changes are risky. Treat each edit as an experiment, not a tweak.
Prompt changes are strategy changes
This is the lesson that reframes everything: a prompt is the strategy. An agent’s instructions determine how it weighs evidence and what it decides, so editing them changes every future call it makes. That deserves the same discipline you’d give a code change — version control, a record of why you changed it, and a way to evaluate whether the change helped.
AlphaFlowSeven versions every per-council prompt edit and can run an LLM auditor (the Council Evaluator) over a council’s recent decisions to propose and assess prompt changes — closing the loop between “we changed the instructions” and “did it actually improve the calls.”
The meta-lesson
Good market-analysis prompting isn’t clever wording. It’s discipline: structured inputs, constrained outputs, sharp roles, and treating the prompt as the strategy it actually is.
Want to see well-engineered prompts in action across six agents? Run a council free — three included, no card.