社区分享需求与可验证测试工作流
LocalLLaMA 一名作者称,其以 6K unsloth 量化运行的 Qwen3.8-27B 在一个周末协助完成了二进制协议构建。作者建议先明确目标、需求、问题和用途,再用 Cucumber 或 EARS 风格文档及可重复测试约束生成过程,并让模型在设计分歧处展示选项。其还称 Claude 审查项目时发现了若干缺失的安全检查,以及 Python 基准脚本中的两处错误。
该帖提供了以需求和可验证测试为核心的具体智能体编程流程,但证据仅来自一名社区用户的单个项目,尚未经过对照评测。
Creating a list of requirements and documentation/tests (in Cucumber or EARS style) with your model is a great way to improve it's code. Just got through a weekend long binary protocol build with Qwen3.8-27B (6K unsloth quant).. and fired up claude to review and it found the foundation was well engineered, but only missed a few security checks and a couple of errors in the python benchmark script comparing the old protocol with the new one.
Really comes down to:
- Design first. Goals, requirements, problems. How is this going to be used.
- Observation, simulation and tests are verifiable vs. assumptions. Tests and processes can be reviewed for "ritual".
- Dialectic: If the model comes to a crossroads, have it present options and NOT just choose-and-go, but initiate a dialog about design.
- Frequent repeatable tests, and review of tests to see if they meet the spirit of the requirement
- Every risk is a requirement.
Here's my prompt set ("skills" I guess) and a tool to install for various local harnesses: https://github.com/bigattichouse/BluePrint