Taming Qwen3 27B's Over-Reasoning With an AI-Generated System Prompt Patch
ea_man · reddit · 2026-09-08
The author hit a wall running Qwen3 dense 27B (Q6XL, KV cache q80, temp 0.85, medium reasoning effort): over-thinking, failing precise instructions, re-evaluating seen data, looping "Actually…/Wait…/Hmm…", meta-reasoning and intent-guessing.
The fix workflow is worth copying:
- Fed old Pi coding session reasoning traces to GPT and Qwen to identify recurring misbehaviors
- Had them generate an APPENDSYSTEM.md patch targeting those pathologies
- Dropped it into /.pi/agent/APPENDSYSTEM.md to steer the agent
Both the patch file and a reasoning pathology report are publicly shared; the author invites others to contribute rules for similar issues.
More from coding & agent
- Shopify design lead on AI-era workflow: designers are becoming full-stack editors — KlausCodes · 2026-10-26
- Ultrametric open-sources Open Startup Repo to let AI agents run startup operations — garrytan · 2026-10-26
- MCP fault injection study: agent task scores drop 60% under realistic tool failures, scale doesn't help — dair_ai · 2026-10-26
- General agents burn ~140M tokens/month on LinkedIn outreach; scripts do it 60x cheaper — ChrisHarpon2 · 2026-10-26
- OpenAI upgrades Decisions API: 400 questions per call, native voice and image support — OpenAIDevs · 2026-10-26
- NLI alone rejects 5/6 new agent memories; a 0.45 similarity gate fixes it all — Alternative_Pin9598 · 2026-10-26