How to Use LLMs in a Trading Bot Stack

Been playing with kraken-cli + Claude Code for a trading bot stack. Here's the split that clicked for me today:

🤖 Agentic for design & validation — describe your strategy in natural language, the LLM figures out the implementation, fetches data, and paper trades. (strategy compiler)

⚙️ Deterministic for execution — once the logic is solid, extract it into a Python/Bash script. Predictable, loggable, no LLM in the hot path. Calculate signals, run strategies. (working implementation)

The LLM reads the script output and reasons over the signal together with related calculated data before asking for human confirmation. (human in the loop)

Idea → code → signal → reasoning → trigger