How Prompt Tuning Improved GPT-5.5 in VS Code

The VS Code team and OpenAI ran a two-week A/B experiment to tune the GPT-5.5 system prompt inside the VS Code agent harness. Two treatments were tested: Treatment A added a compact 'economical search and edit' reminder, while Treatment B restructured the prompt into explicit before/after-first-edit sections. Treatment B won decisively — reducing average tool calls by 8.54%, cutting p95 tail-end token usage by 7.64%, and speeding up median time-to-first-edit by 5.68% and p95 latency by 9.30%, all with highly significant p-values. Quality guardrails (code survival rates) remained largely stable. Treatment B is now the default GPT-5.5 system prompt in VS Code, and the team plans to continue this optimization loop across models and harness configurations.