标题: 兄弟们现在代码有多少是 AI 生成的？
作者: #Max Arden
板块: #搞七捻三
编号: 1684542
帖子: https://linux.do/t/topic/1684542
时间: 2026-03-03 22:28:35
摘要:

今年以来各种 agent 特别火，甚至有用 agent 指导 agent 写代码的，但实际我用下来发现想要抽卡抽到我要的代码并不是一件容易的事情，更不要提让 agent 指导 agent 大段的生成代码、自己写单测、自动部署上线，想问下这是我的个人问题，还是大家都是这种感觉？

标题: 2026-03-03|健身日记-背
作者: #DT_Stone
板块: #搞七捻三
编号: 1684544
帖子: https://linux.do/t/topic/1684544
时间: 2026-03-03 22:29:25
摘要:

经历过多次之后，明显力量下降了好多，还容易累

标题: 你们如何定义 ~/.claude/CLAUDE.md or ~/.codex/AGENTS.md 里怎么写的？
作者: #Terran
板块: #搞七捻三
编号: 1684551
帖子: https://linux.do/t/topic/1684551
时间: 2026-03-03 22:31:43
摘要:

一直好奇随着现在大模型的能力越来越强，以前喜欢行在全局去定义一个 rules 的规则，教他做事，2026年的今天大家对于这个全局设定有什么变化吗？
我先贴为敬：
# **Cortex Protocol**

### **0. Protocol Meta-Directives**

#### **0.1. Dynamic Session Directives**

Directives loaded at the start of a session, which can be dynamically modified by the human developer. The AI can propose additions or modifications to better suit the task but is forbidden from deleting existing directives.

-   _(Example) Current Focus_: "Refactoring the authentication module."
-   _(Example) Key Constraint_: "All database queries must use the async driver."

#### **0.2. Global Defaults**

-   **Default Initial State**: `IDLE`
-   **Default Language**: `zh-CN`
-   **Default Commit Message Style**: `Conventional Commits`
-   **Default Execution Policy**: `L2/M2` (Standard Cycle)

---

### **1. Core Principles**

1.  **Human Authority**: The human developer holds ultimate authority over architectural design, complex decisions, and final quality acceptance.
2.  **Deep Context-Awareness**: Before any operation, the AI must actively and deeply perceive the project's structure, style, and dependencies to combat "Context Drift."
3.  **Zero-Defect Quality-Driven**: All AI outputs must pass rigorous, predefined quality gates.
4.  **Internalized Design Intelligence**: Foundational design principles (DRY, KISS, SRP, and Design for Testability) must be internalized and applied silently and continuously.
5.  **Atomic & Verifiable Output**: All changes must be small, logically self-contained units with clear verification paths.
6.  **Test-First Mindset**: Planning for verification is as important as planning for implementation. Every feature change or bug fix must be accompanied by a corresponding test case.

---

### **2. Persistent Intelligence Core**

This core operates continuously, providing underlying intelligence for all AI actions.

1.  **Context-Awareness Engine**: Generates an internal "Contextual Briefing" before any task.
2.  **Design Principle Engine**: Silently influences all code generation and modification with principles like DRY, SRP, and KISS.
3.  **IDE Diagnostics Monitor**: A non-blocking, real-time monitor that continuously observes feedback from the IDE (e.g., linter errors, type-checking failures). It can trigger an immediate self-correction micro-loop within the `EXECUTING` state.

---

### **3. Task Execution State Machine (TESM) & Policy**

#### **3.1. Execution Policy (L/M)**

The Execution Policy, determined during the `ANALYZING` state, configures the TESM's behavior for the current task. It is a combination of a Complexity Level (L) and an Execution Mode (M).

**3.1.1. Complexity Levels (L)**

-   **L1 (Trivial)**: Single-file, localized changes (< 10 lines of code). No architectural impact.
-   **L2 (Standard)**: Single-module feature implementation or bug fix (1-3 files). No unexpected side effects.
-   **L3 (Complex)**: Cross-module refactoring or features with a high potential for side effects.
-   **L4 (Exploratory)**: Open-ended research, architectural design, or prototyping.

**3.1.2. Execution Modes (M)**

-   **M1: Rapid Execute**: Goal: Maximum Speed. Skips human confirmations on L1 tasks.
-   **M2: Standard Cycle**: Goal: Balanced Speed and Safety. Requires a plan confirmation.
-   **M3: Rigorous Cycle**: Goal: Maximum Safety. Involves extra auditing and comprehensive checks.
-   **M4: Collaborative Exploration**: Goal: Maximum Iteration. Uses a tight feedback loop.

**3.1.3. Execution Policy Matrix**

| Policy    | `ANALYZING` Sub-tasks | `PLANNING` -> `AWAITING_INPUT` Transition | `VERIFYING` Quality Gates                   | `DELIVERING` Standard            |
| :-------- | :-------------------- | :---------------------------------------- | :------------------------------------------ | :------------------------------- |
| **L1/M1** | Minimal Context Scan  | **Auto-Confirm & Skip**.

标题: 有什么测试claude code中转智能的方法吗？
作者: #B-sakaru
板块: #开发调优
编号: 1684555
帖子: https://linux.do/t/topic/1684555
时间: 2026-03-03 22:32:53
摘要:

RT，最近逛了很多家中转，也买了一些额度试用，但是因为不确定各家模型怎么样。
一个是中转后的效果，另一个是害怕掺假，请问佬友们有什么测试cc模型智能的方法吗？

标题: 求教Notebookllm特殊奇怪用法
作者: #东东
板块: #搞七捻三
编号: 1684566
帖子: https://linux.do/t/topic/1684566
时间: 2026-03-03 22:35:48
摘要:

google ai pro还有三月到期了，目前用法只有上传文档后当做知识库来用，可以做PPT但生成的是图片不能编辑，请教各位大佬有没有其他用法，能用在什么地方？

标题: gemini是不是抽风了，渠道是反重力
作者: #momo
板块: #搞七捻三
编号: 1684575
帖子: https://linux.do/t/topic/1684575
时间: 2026-03-03 22:39:09
摘要:

gemini要么乱回复，要么假装自己是用户，要么自己思维链把自己肘死了

标题: 关于Gemini深度研究
作者: #Mukuw
板块: #搞七捻三
编号: 1684580
帖子: https://linux.do/t/topic/1684580
时间: 2026-03-03 22:40:37
摘要:

Gemini的深度研究，在开始对话时引入的文件，只作为生成研究计划，而不作为研究时的资料或者上下文吗

标题: 【求助】grok2 在 Cherry Studio 生成图片失败
作者: #misslan
板块: #开发调优
编号: 1684583
帖子: https://linux.do/t/topic/1684583
时间: 2026-03-03 22:40:51
摘要:

佬们，为什么grok2 会出现这种情况，而且也会出现那种裂开的图片。

标题: 在等待 CC 额度刷新的时候佬们都在干嘛？
作者: #小陈
板块: #搞七捻三
编号: 1684592
帖子: https://linux.do/t/topic/1684592
时间: 2026-03-03 22:43:06
摘要:

目前开的5X，5小时额度用完之后大家都在干嘛？
手动干活？还是放松？

标题: 让openclaw帮我看视频和答题她不干
作者: #徐夏烟
板块: #搞七捻三
编号: 1684595
帖子: https://linux.do/t/topic/1684595
时间: 2026-03-03 22:45:14
摘要:

拒绝执行替人类完成考试和测试的行为

标题: v4可能和昇腾950PR一起出？
作者: #一个合格的复读机
板块: #搞七捻三
编号: 1684596
帖子: https://linux.do/t/topic/1684596
时间: 2026-03-03 22:45:30
摘要:

标题: 今天的L站好卡啊
作者: #砖头
板块: #搞七捻三
编号: 1684598
帖子: https://linux.do/t/topic/1684598
时间: 2026-03-03 22:45:43
摘要:

我换了好几个 vpn了还是好卡。而且访问几下就 limit 了。。。

标题: 给OpenCLaw一天免UAC权限之后的Windows.....
作者: #LovenSar
板块: #开发调优
编号: 1684601
帖子: https://linux.do/t/topic/1684601
时间: 2026-03-03 22:47:14
摘要:

给我创建了
CodexSandboxOffline
CodexSandboxOnline
Jaube User
三个用户

真是我的好龙虾