Cline把Meta的Muse提示词搬到自己框架里,修bug又快又便宜,值得看看他们怎么写的。
Meta发布Coding Agent Muse Code和模型Muse Spark 1.2,Cline团队因登录bug改用提示词实验。他们从Muse Code系统提示词中提取五条工作纪律,移植到Cline的harness中。同一模型Muse Spark 1.2修复同一bug,Token消耗从19.7M降至7.2M,耗时从49分钟减至24分钟,成本从7.69美元降到3.25美元。实验证明提示词对智能体性能影响显著。
Muse Code (beta) 和 Muse Spark 1.2 有趣的实验 Meta 同时发布了这对 Coding Agent 和 LLM,@cline 团队用它做了个有趣的实验,顺便提取了系统...
Muse Code (beta) 和 Muse Spark 1.2 有趣的实验 Meta 同时发布了这对 Coding Agent 和 LLM, @cline 团队用它做了个有趣的实验,顺便提取了系统提示词。 Cline 团队原本想直接试用 Muse Code 智能体,但发现它存在一个 bug——无法在 Docker 容器中登录,于是改做了一个更有意思的实验。 Meta 官方称 Muse Spark 1.2 与其 Muse Code Harness 是协同训练的,即模型和框架共享同一套指令格式,模型在训练时见过框架的轨迹数据。Cline 团队由此想到:如果把 Muse Code 系统提示词中的指令提取出来,移植到 Cline 自己的 harness 里,会发生什么? 提取出的五条提示词原则 Cline 给出了这套"特殊提示词"的核心内容,本质上是一套严谨的工作纪律: · 信任源代码胜过信任用户提示 —— 开始任务前先读完每一个调用点和已有测试; · 边界情况和错误情况与主路径(happy path)同等重要; · 修复 bug 之前,必须先复现它; · 不要相信第一次全绿的测试套件 —— 对看起来敷衍、半成品的测试要核实; · 不要改完就停 —— 持续工作直到改动被验证真正完成。 这五条都指向同一个倾向:对抗"急于收工"的本能,把验证和尽调放到与写代码同等重要的位置。 实验结果 他们用改造后的 Cline harness 修复一个真实的仓库 bug,与原版 Cline harness 对比。同一模型(Muse Spark 1.2)、同一任务,只有提示词不同: 原版 Cline vs. 移植 Muse 提示词后 Token 消耗:19.7M vs. 7.2M -- 少 2.7 倍 耗时:49 分钟 vs. 24 分钟 -- 快 2 倍 成本:$7.69 vs. $3.25 -- 便宜 2.4 倍 Cline @cline We tried using Meta's new Muse Code agent, but it has a bug that doesn't let it sign in from a docker container. So we did a fun experiment: Meta claims Muse Spark 1.2 was co-trained with their Muse agent harness. So we extracted instructions from their system prompt and added them to the Cline harness. TL;DR of this special prompting: - Trust source code over the user prompt, so read every call site and existing tests before starting the task - Weigh edge and error cases as heavily as the happy path - Always reproduce the bug before fixing - Don't trust the first passing test suite, and verify suspicious looking half-baked tests - Never stop at just editing, keep working until the change is verified complete. We then asked this modified harness to fix a real bug from our repo, and compared the results to the original Cline agent harness. Results: - Used 2.7x fewer tokens (19.7M → 7.2M) - Finished 2x faster (49min → 24min) - Cost 2.4x less ($7.69 → $3.25) Same Muse Spark 1.2 model, same task, only the prompting changed. Incredible how much of a performance gain Meta was able to achieve training it on these special instructions! 🔗 View Quoted Tweet 💬 0 🔄 0 ❤️ 4 👀 302 📊 1 ⚡