1. Role separation

Define explicit responsibilities (Analyst = quant report, Boss = directive, Executor = PnL + execution guard) and keep prompts < 4k tokens.

2. Cooldowns & budgets

Use `max_trades_per_hour`, `cooldown_minutes`, and per-role model choices (flagship models for analysis, faster models for executor) to prevent runaway token spend.

3. Semantic stops

Embed instructions like “Reject execution if edge < cost” or “Skip trade if missing market data” to keep the LLM from improvising.

4. Versioning & rollout

Prompts live in `system_prompts.json`, versioned in session logs, and rolled out through Brain Control Center with rollback.