Swapping just the system prompt measurably changes harness benchmark results
yb2698 · x · 2026-10-07
The author continues their harness-impact experiments by comparing two ways of changing the system prompt:
- v1: appends an extra instruction to the existing Pi system prompt
- v2: replaces the default prompt with a single minimal sentence
They also shipped a Hugging Face Space demo where you can swap the system prompt and tools and watch the model attempt a terminal-bench 2.1 task live.
More from coding & agent
- Gesture recognition with OpenAI's vision Decisions API beats earlier demo — pvncher · 2026-10-07
- magpie: open-source AI agent infra devs feed to their agents as a reference codebase — vista8 · 2026-10-07
- magpie: open-source menu bar hub to run Codex on DeepSeek, Claude Code on Kimi — vista8 · 2026-10-07
- Hugging Face Transformers 5.19.0 ships multiple breaking changes; run your evals before upgrading — lmoroney · 2026-10-07
- Giving away your Legos—again, this time to a team of agents — danshipper · 2026-10-07
- AI didn't just lower the bar—it raised the ceiling of what one person can build — jonathan_wilke · 2026-10-07