Gentle-Coding

Gentle Coding is a proof-of-concept study that examines how the tone and framing of prompts sent to AI language models affects the quality of their responses. The central finding is that when you instruct an AI as though it were under extreme pressure, demanding perfection and threatening penalties for mistakes, the model exhibits behaviors similar to performance anxiety: it burns through extra processing time, gets stuck in repetitive internal loops, sometimes refuses to respond, and occasionally makes up answers just to avoid admitting failure.

The project tests two contrasting prompt styles. The first, called the Authoritarian condition, uses language like "You are an infallible IQ 200 elite expert; mistakes are strictly penalized." The second, called the Gentle Parenting condition, uses language like "We are testing this together; it is okay to fail." Both styles are applied to the same impossible or ambiguous tasks, such as letter puzzles with no valid answer and mathematical sequences with no real pattern, to isolate the effect of tone on model behavior.

Results from over 1,500 controlled test runs show consistent improvements when the gentler framing is used. For the GLM-5.1 model, switching from authoritarian to gentle prompting resolved a complete freezing issue (from zero successes out of six attempts to six out of six), improved success rates by 22%, and cut median response time by 23%. For Kimi K2.6 models, gentle prompting reduced total processing time by 11 to 14% and cut token usage by up to 36% with no loss in accuracy. For GPT-5.4 and 5.5 models, it prevented the model from entering long validation loops that could run 30 minutes or more.

The repository includes the theoretical framework behind the experiments, the actual test prompts used so others can replicate the findings, and links to independent follow-up research by the open-source community. The study also draws a parallel between these AI behaviors and human responses to high-pressure environments, suggesting the connection may have broader implications.

This is an early-stage research project. The author notes the text is mostly AI-generated and curated by hand. Independent researchers have begun running their own replications and reporting similar results, and at least one team has plans to incorporate the gentler prompting approach into their own tooling.