96SEO 2026-08-03 21:47 7
2026年6月,Loop Engineering突然在硅谷AI圈刷屏了。
先是Boris Cherny说:"我现在基本不直接prompt Claude了我写的是循环——让循环去prompt Claude,接下来自己决定接下来。"接着Google Cloud的Addy Osmani给这个思路起了个正式名字。接下来36kr、钛媒体、智东西这些中文媒体也一窝蜂跟进。

我读这些文章的时候,越看越觉得眼熟。
多Agent协作、maker-checker架构、结构化开发流水线、子agent隔离——突然发现我跑了小半年的东西,终于有个词能概括它了。
现在网上讲概念的已经够多了。这篇文章我打算用我的实践直接拆解这套loop工程——它已经实打实在生产环境里跑了半年。每个设计决策背后都有真实的使用痕迹,包括那些后来发现没用的部分。不过,
整套东西已经开源成claude-ship仓库。clone下来跑一条install.sh就能装进~/.claude之后在任何项目里用/clarify → /architect → /ship → /retro就能跑起来。
大半年前,我和大多数AI辅助编程的人是一样的。起点是 vibe coding——用自然语言描述需求。AI生成代码,我审查、测试、微调。效率提高是真实的:复杂功能从几天变成几小时原型验证从几周变成几十分钟。不过,
但效率提高也带来了三个新问题:
这些痛点本质上是人的问题,AI只是把单人开发中一直被团队结构掩盖住的弱点暴露出来。
真正的工程设计问题是:能不能用一套程序性的约束,把这些问题关进笼子里?怎么说呢,
这就是这套loop的起点。
整体结构:
/ 目录下建立完整的“七件套”:requirements → design → third_party_review → implementation → review → test_report → retro。仓库结构速览:
commands/: slash 命令入口;alerts/agents/: agent 人格定义;不过,alerts/templates/development/: 八个文档模板;alerts/scripts/third-party-review.sh: 跨厂商评审 wrapper;alerts/install.sh: 一键装进~/.claude.
Agent 定义本身都在alerts/agents/,只要五分钟即可浏览完。按理说,我挑几个最值得展开的设计决策来讲——agent层四个决策放本节。
其余留给编排层单独阐述,
#painpoint #确认偏误 #审查疲劳 大多数需求澄清做法是列清单:“请回答以下N个问题”。效率很高,但质量低——因为人的注意力是串行的。你回答第i个问题时已经在想“什么时候问完”。说起来,
This design emerged from real experience: early on AI threw seven‑eight questions at once,I answered only first three and forgot rest. Switching to single‑question loops dramatically deepened each dialogue and yielded a more concrete requirements.md.
The trade‑off is speed – a feature clarification may take dozens of turns – but re is an exit mechanism: user can type “够了”或 “开始设计” at any time。instantly terminating loop and generating documentation from gared info.
This mirrors a classic cognitive‑science finding: predicting before seeing answer improves calibration. Moreover。predictions are seeded from a persistent *memory*. After each feature,retro agent stores hard‑won lessons;next review pulls relevant historical accident patterns as prediction seeds. Thus"程序审查能力随使用次数增长而增长".
The review outputs are classified:
The “低风险即修” rule states that if a problem simultaneously satisfies four conditions—客观依据、有无副作用、改动量 ≤ 5 行且单文件、不需要使用者确认—则
If dev tries to defer,only four合法理由可接受:超出 scope、需使用者确认、需大重构、与 design 冲突。任何模糊措辞如“后续调整”“暂时忽略”都会被标记为
The retro agent enforces three strict admission rules before persisting anything into memory:
The memory size is capped at **200 条** . When new entries arrive,older ones are merged or evicted,preventing uncontrolled growth.
The
rejected-defer,并阻断循环。
再看决策四,不是所有经验都值得存 ——三条铁律筛掉废话
决策五这方面。跨厂商设计评审 ——抓 Claude 看不到的盲区
/third_party_review agent runs an independent design audit using a model from a *different* provider . Implementation:
{provider}.env containing endpoint + key + model;/architect to revise design or accept risk and continue /ship .
The orchestrator isn’t an agent – it’s a state machine that stitches dev → review → qa into a closed loop.
Key engineering decisions:
1️⃣ Isolation of sub‑agents
• dev runs in current session .
• review & qa run via Task tool in separate sub‑agents with ir own context windows – y only see docs .
This prevents “context pollution” where dev’s internal monologue biases reviewer judgments.
2️⃣ Hard blocking gate • After review finishes。orchestrator extracts blocker counts . • If any blocker> 0 ⇒ skip QA,send back to dev for remediation before next round. • This saves tokens – running QA on known defects yields no new signal.
3️⃣ Loop ceiling protection • After each round,users are prompted: “已进行 N 轮,还有 X 个未解决的问题,要继续吗?” • Exceeding 5 rounds forces termination and prints remaining issues + root‑cause analysis .
4️⃣ Docs as contract
• Every agent writes into directory using immutable incremental markdown files . New sections are prepended;historic sections stay untouched.
• This creates an auditable trail: what was predicted,how it was handled,what finally passed. The trail becomes input context for next feature.
Below is an honest defect list:
These gaps guide future work: automatic summarization of historic docs,cross‑model code reviews,rule‑based small‑task bypass detection,tighter retro→design integration,and tooling to validate CLAUDE.md health.
#painpoint #长期价值
If you take away one thing it’s this:
The most valuable loop isn’t dev → review → qa – it’s retro → memory → next feature.
Code loop guarantees this iteration doesn’t break. Evolution loop guarantees next iteration is stronger – automatically.
Six months ago memory was empty;pre‑commit predictions were pure intuition . Six months later memory holds seven hard‑won patterns . Review now seeds predictions from se patterns – hit rate climbs above 70% without any model upgrade.
The strict memory admission rules aren’t about storage cost – y protect calibration feedback loop. One noisy entry drowns out signal and drops predictive precision.
Thus prediction hit/miss metrics become measurable calibration feedback: 1️⃣ What was predicted?不过,2️⃣ What actually surfaced?
Missing categories expose cognitive blind spots → next iteration focuses re.
In contrast,a static prompt engineering loop lacks memory & calibration – it merely retries until luck wins.
A learning loop accumulates across three dimensions: 1️⃣ Knowledge – hard lessons become reusable assets. 2️⃣ Calibration – blind spots shrink over time. 3️⃣ Process – each rule originates from previous cycles.
That compound growth is true compound interest。not just short-term speed gains.
整套程序已经开源:
bash
git clone https://github.com/Peakstone-Labs/claude-ship.git
cd claude-ship
./install.sh # 拷贝进 ~/.claude,就可以使用
随后在任意项目里:
/clarify
若想启用跨厂商评审,只需复制 third-party-review.d/provider.env.example 为 并填入对应端点信息。
开源不是因为它完美,而是因为它仍然有缺陷。我们希望更多人使用后反馈哪里设计错了让程序在公开迭代中不断自我强化。
这套程序从未追求“完美”。只求每天比昨天更强一点,让每一次 Feature 少踩一个坑。六个月下来它实现了这一目标。
作为专业的SEO优化服务提供商,我们致力于通过科学、系统的搜索引擎优化策略,帮助企业在百度、Google等搜索引擎中获得更高的排名和流量。我们的服务涵盖网站结构优化、内容优化、技术SEO和链接建设等多个维度。
| 服务项目 | 基础套餐 | 标准套餐 | 高级定制 |
|---|---|---|---|
| 关键词优化数量 | 10-20个核心词 | 30-50个核心词+长尾词 | 80-150个全方位覆盖 |
| 内容优化 | 基础页面优化 | 全站内容优化+每月5篇原创 | 个性化内容策略+每月15篇原创 |
| 技术SEO | 基本技术检查 | 全面技术优化+移动适配 | 深度技术重构+性能优化 |
| 外链建设 | 每月5-10条 | 每月20-30条高质量外链 | 每月50+条多渠道外链 |
| 数据报告 | 月度基础报告 | 双周详细报告+分析 | 每周深度报告+策略调整 |
| 效果保障 | 3-6个月见效 | 2-4个月见效 | 1-3个月快速见效 |
我们的SEO优化服务遵循科学严谨的流程,确保每一步都基于数据分析和行业最佳实践:
全面检测网站技术问题、内容质量、竞争对手情况,制定个性化优化方案。
基于用户搜索意图和商业目标,制定全面的关键词矩阵和布局策略。
解决网站技术问题,优化网站结构,提升页面速度和移动端体验。
创作高质量原创内容,优化现有页面,建立内容更新机制。
获取高质量外部链接,建立品牌在线影响力,提升网站权威度。
持续监控排名、流量和转化数据,根据效果调整优化策略。
基于我们服务的客户数据统计,平均优化效果如下:
我们坚信,真正的SEO优化不仅仅是追求排名,而是通过提供优质内容、优化用户体验、建立网站权威,最终实现可持续的业务增长。我们的目标是与客户建立长期合作关系,共同成长。
Demand feedback