The Surprising Prompt Technique That Gives Gemini Reasoning a Massive Boost

Bruce Middleton’s CoT-forcing promptware, shared on LessWrong, is a novel method to enhance AI reasoning by embedding structured rules within a system directive. The core tree_rule forces the model to list N important perceptions at each phase (e.g., phase 1: N perceptions; phase 2: N perceptions from each of those), then affective effects and elicited behavior. This creates a tree-like chain-of-thought. Additional rules include a halt_rule that aborts if instructions are problematic, a length_rule to limit output sentences, and a check_rule to verify context presence. The prompt_rule terminates generation immediately after resolving primary input, adding a "Standing by" marker, which also acts as a dead-man switch if FIFO truncation removes the directives.

Practical tests with Gemini showed the technique effectively eliminates distracting question prompts by satisfying the model’s need for a terminal marker. The author reports that even a self-contradictory prompt triggered the halt_rule once. By placing the prompt_rule at the top, it serves as a safety net: if the context window truncates the rules block, a non-standard prompt is issued. This approach allows fine-grained control over reasoning depth and output length, making it valuable for researchers and developers needing reliable, distraction-free model responses in agent modeling tasks.