Steer Models Mid-Reasoning to Save Time and Tokens
Correct the AI's logic while it is still reasoning.
Use the mid-reasoning instruction feature in GPT-5.5 Thinking to steer the model's logic before it generates a final response.
The Scenario
You are using GPT-5.5 Thinking for a complex data analysis and notice in the preamble that it plans to use a methodology you didn't intend. You need to adjust its course before it generates the full, incorrect report.
Before & after
In older models, you had to wait for the entire response to finish before providing feedback, often requiring multiple 10-minute cycles of trial and error.
By adding follow-up instructions while the reasoning preamble is visible, you can pivot the model's logic instantly. This takes 1–2 minutes to ensure the output aligns with your requirements.
The Prompt
[DURING_THINKING_PHASE] Actually, I need the analysis to focus specifically on the [INSERT_KEY_METRIC] instead of the general overview you planned. Adjust your reasoning accordingly.
GPT-5.5 Thinking and Pro models now allow for real-time steering. When the 'Thinking' process begins, the model provides a preamble of its plan; users can interrupt or add constraints immediately rather than waiting for a completed response that might have missed the mark.
Source
Model Release Notes | OpenAI Help Center"You can add instructions while the model is still thinking to steer the response before it finishes."