Claude 代码团队应该尝试宏功能,以便用户能完成三倍的任务量。
1 分•作者: yohji1984•3 个月前
所以这个想法非常简单。如果 CC 需要更改一个文件并运行一些测试,CC 需要:
# 回合 1 — 应用补丁(更改包文件)
# 回合 2 — 应用补丁(修复 bug)
# 回合 3 — 应用补丁(编辑测试脚本)
# 回合 4 — 构建
# 回合 5 — 运行测试和 lint
# 回合 6 — *如果使用 playwright,CC 需要再进行一轮读取媒体
概念是我们可以使用宏命令将整个过程在一个 RAG 中,在一个回合内完成。例如像这样:
step1: 检查环境
step2: 应用补丁(更改包文件)
step2: 应用补丁(修复 bug)
step2: 应用补丁(编辑测试脚本)
step3: 构建
step4: 运行测试和 lint
step5: 读取媒体
由于 Antropic 的使用条款,我无法测试此实现。但我曾在其他编码代理和提供商上测试过,这可以将 LLM 的回合数减少 40% - 80%,并节省近乎相同的 token 量。(我需要使用 ** 来隐藏提供商的名称,以避免帖子被机器人删除)。
如果您认为这种方法有帮助,并且我的测试方法是合理的,您可以给我点赞!谢谢!
实现和测试基准可以在 md 文件中找到。我使用 DeepSWE 和整个仓库重写案例进行了测试:
https://github.com/Tura-AI/tura
查看原文
So the idea is really simple. If CC need to change a file and run some test, CC needs to:
# Turn 1 — apply patch (change package file)
# Turn 2 — apply patch (fix the bug)
# Turn 3 — apply patch (edit the testing script)
# Turn 4 — build
# Turn 5 — run tests and lint
# Turn 6 — *and if is playwright CC needs another round to read media<p>The concept is that we can use macro command to put the entire process into a RAG in one single turn. For example like this:<p>step1: inspect env
step2: apply patch (change package file)
step2: apply patch (fix the bug)
step2: apply patch (edit the testing script)
step3: build
step4: run tests and lint
step5: read medias<p>I could not test this implementation due to Antropic terms of use. But I tested on other coding agent and providers this can reduce llm turns around by 40% - 80%, and saving nearly same amount of tokens. (I need to use ** to hide the provider's name in order to avoid the post being deleated by the bot).<p>If you think this approch could help and my testing method is reasonable, you can give me an upvote! Thx<p>Here is the implementation and testing benchmark can be found in the md. I tested with DeepSWE and entire repo rewrite cases:<p>https://github.com/Tura-AI/tura