Recursive prompt engineering: how Gemini 3.0 Pro refined and executed its stock-picking playbook
We made Gemini 3.0 pro re-engineer its own deep research workflow using results from the last run.
📶BUY $TXT (Textron Inc) and hold for 22 months till November 19th, 2027.
On January 4, 2026, we tasked Gemini with designing a workflow to leverage its Deep Research capabilities. The goal was to identify "information asymmetry" in the stock market and generate a high-conviction recommendation. The resulting pick, Parsons Corporation ($PSN), has already returned ~16.4% in just 18 days.
For this iteration, we are executing a technique called Recursive Prompt Engineering. We fed the original prompt structure back into Gemini 3.0 Pro, along with the performance data from the first run, and asked the model to self-optimize the strategy for even higher returns.
Crucially, we also removed the arbitrary one-year holding constraint. Instead, we instructed the model to dynamically determine the optimal exit date to maximize the velocity of capital.