导语：从“提示词炼金术”到“工程理性”
过去两年，整个技术圈几乎陷入了一种狂热的“提示词炼金术”。
每个人的剪贴板里都存着几套长达数千字的 System Prompt，充斥着密密麻麻的条条框框：
“你是一个世界顶级的资深架构师……你必须严格遵守 Clean Code 原则……绝对不能省略代码……在输出之前请一步一步深度思考……如果你表现好我会给你 200 美元小费……”
我们试图用这些近乎玄学的自然语言“咒语”，去规训一个参数量庞大的黑盒。
但真实的工程结果往往极其残酷：提示词越堆越长，模型的注意力越涣散；原本在单点上表现良好的指令，在大上下文里频繁被稀释。写 Prompt 越来越像是在向神明祈祷，运气成分远大于科学。
最近几周，GitHub 趋势榜发生了一场极具里程碑意义的剧变：由开源老兵 Jesse Vincent 发起的 obra/superpowers 席卷全网，同时 Cloudflare、NVIDIA、AWS 等云巨头纷纷下场开源官方的 Agent Skills。
这场集体转向释放了一个无比清晰的信号：靠自然语言堆砌的“提示词工程”正在加速破产，而经典的“软件工程纪律”，正在以 Agent Skills（智能体技能化）的形态全面复兴。
一、 提示词的极限：口头叮嘱从来不是制度保障
为什么手写几千行 Prompt 无法从根本上解决 AI 乱改代码、逻辑断片的毛病？
因为**“口头叮嘱”在系统工程里是最低效的管理手段**。
这就像团队里招来了一位博览群书但经验极度缺乏的天才新人。你在他耳边唠叨一千遍“记得写单元测试、不要破坏原有接口、不要随意改动公共配置”，一旦面对复杂的几万行真实业务，他依然会在一瞬间把核心逻辑改得面目全非。
传统的 Prompt 交互存在三大致命死穴：
- 注意力稀释（Attention Dilution）：把几千字的规范一股脑塞给模型，模型会在漫长的前向计算中逐步淡忘早期的约束条件；
- 缺乏状态机与确定性阻断：Prompt 无法强制约束步骤。你让模型“先写测试再写实现”，模型经常在一次回答里把测试和实现全写了，根本没有验证测试是否真的能拦截 Bug；
- 脆弱的不可维护性：Prompt 是脆弱的散装文本。模型底层微调一次，或者微调参数稍有变动，原本生效的 Prompt 就会莫名失效，根本无法纳入严肃的团队版本控制。
二、 Agent Skills 的本质：给野马套上“确定性缰绳”
以 superpowers 为代表的技能化框架，到底做对了什么？
它的核心转变在于：它不再试图去“教”模型怎么写代码，而是给模型打造了一套“无法越界的轨道”。
1. 强制执行硬核的工程方法论（如 TDD）
在 superpowers 的技能库中，最核心的一项能力是测试驱动开发（TDD）规范。它不是一句空洞的“请先写测试”，而是把工程纪律固化为确定性的行动契约：
- 第一步：强制要求智能体编写一个能够精准复现 Bug 的失败测试（红灯）；
- 执行拦截：在测试真实运行且失败之前，系统的工具权限直接锁定，禁止修改任何生产业务代码；
- 第二步：仅允许编写能让测试转绿的最小代码量；
- 第三步：自动运行全量回归，确保原有链路零破坏。
在这套流水线下，AI 不再是一个信马由缰的即兴演说家，而是一个被按在工业流水线上的精密执行器。
2. 上下文按需挂载，告别“全家桶污染”
在传统模式下，为了让 AI 懂规范，我们往往要在会话开篇投喂大量的架构规范、设计模式、Lint 规则。
而 Agent Skills 采用了极度克制的动态挂载机制：
平时，核心上下文极其干净，只有最基础的任务目标；只有当 Agent 决策“当前需要进入安全漏洞排查”时，系统才会动态挂载对应的 security-audit.skill；一旦审计完成，相关上下文立刻卸载。
用即加载，用完即弃。 这不仅为长程任务省下了高达 70% 的昂贵 Token，更保证了核心推理链条始终维持在高信噪比的状态。
3. 将玄学沉淀为“代码化资产”
Prompt 是给模型读的散文，而 Skill 是结构化的工程资产（标准化清单、触发条件、检验钩子、辅助脚本）。 这意味着团队可以像管理生产代码一样，对每一个 Agent Skill 进行 Git 审查、Diff 比对、持续集成和灰度发布。软件工程的确定性，终于重新接管了 AI 的不确定性。
三、 历史的回响：古老的工程纪律，成了大模型时代的救生索
计算机科学发展了半个多世纪，人类前仆后继总结出来的那些经典法则——测试驱动、单一职责、状态隔离、防御性编程，最初发明出来的目的，从来不是为了讨好计算机，而是为了防范人类程序员自身的疏漏与混乱。
极其戏剧性的是，五十年后的今天，当我们面对一个计算速度快上万倍、但也更容易产生幻觉的大模型时，最能驯服它的武器，依然是这套古老的工程智慧。
大模型不需要你在 Prompt 里用尽修辞去教它 Python 怎么写，它已经把全世界的开源代码背得滚瓜烂熟。
它缺少的不是智商，而是戒律。
从“手写几千字 Prompt 碰运气”，到“用标准化的 Agent Skills 建立工程流水线”，这场正在 GitHub 上演的潮流，不仅是开发工具的迭代，更是 AI 编程正在真正走向成熟的成年礼。
💡 关注【一缕82年的清风】，洞悉技术底层与生态演进
欢迎在评论区探讨交流与点赞转发
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Introduction: from “prompt alchemy” to “engineering rationality”
Over the past two years, the entire technology community has been mired in a fanatical “prompt alchemy”.
Each of them has several sets of System Prompts with thousands of words in their clipboard, filled with insulated boxes:
"You are the world's top senior architect... you must strictly adhere to the Clean Code principles... and never omit the code. Think in depth one step at a time... and I'll give you $200 tip if you're good."
We're trying to teach a big black box of parameter count in these near-intensity natural languages.
But the real work results are often extremely cruel: the longer the prompt pile, the less attention the model is paying; the better-performing instructions on a single point are frequently diluted in large context. Writing Prompt is more and more like praying to God, and luck is much more important than science.
In recent weeks, GitHub's trend list has undergone an extremely landmark change: the obra/subperpowers, initiated by Open Source Veterans Jesse Vincent, have swept the web, while the official Agent Skills, like Cloudflare, NVIDIA, AWS, and others, have fallen.
The collective shift has sent a clear signal that the “prompt Project”, built on natural languages, is accelerating bankruptcy, while the classic “software engineering discipline” is fully rejuvenating in the form of Agent Skills (AI agent).
I. Limits of prompt: Oral advice has never been an institutional guarantee
Why is handwritten thousands of lines Prompt not able to solve the root causes of AI's altercation of code, logical fragmentation?
Because** “oral dingling” is the least effective management tool in system engineering**.
It's like the team's got a new talent who's been reading a collection of books, but very inexperienced. You've been whispering to him a thousand times, "Remember to write cell tests, not to break the original interface, not to change the public configuration at will," and once he faces a complex, tens of thousands of lines of real business, he'll still be turning the core logic into an instant.
The traditional Prompt interacts with three deadly graves:
- Attention Dilution: a few thousand words of discipline is plugged into a model, which gradually fades from earlier constraints in long forward calculations;
- Lack of a status machine and certainty block: Prompt cannot impose binding steps. You get the model to "write the test before write it" and the model often completes it in one response, and there is no proof that the test actually stopped Bug;
- Vulnerability: Prompt is a fragile bulk text. The model is fine-tuned once at the bottom, or slightly fine-tuned the parameters, so that the originally valid Prompt will lapse and cannot be incorporated into serious team version control.
II. The Nature of Agent Skills: Placing the Mustang with a “certainty cord”
What is the right thing to do with the skills framework represented by superpowers?
Its core shift is that it no longer attempts to “teach” the model how to write the code, but rather to create a “untransboundary orbit” for the model.
Enforcement of hard core engineering methodologies (e.g. TDD)
One of the core competencies in the skills base of superpowers is the test-driven development (TDD) norm. It is not an empty “preliminary test”, but it is a contract of action that solidifies engineering discipline into certainty:
- First step: oblige AI agent to prepare a failed test (red light) capable of accurately reproducing Bug;
- Executing interception: the system ' s tool permission is locked directly before testing the real operation and failure, prohibiting any modification of the production business code;
- Step 2: Only the minimum number of codes allowed to turn the test green;
- Step three: Auto-run full-scale return to ensure zero damage to existing links.
Under the current line, AI is no longer a improvist, but a sophisticated implementer that is pressed on the industrial current line.
context mounted on demand to say goodbye to “bunk contamination”
In the traditional mode, in order for AI to understand the norms, we tend to invest in a lot of architecture, design models, Lint rules at the beginning of the session.
Agent Skills uses an extremely restrained dynamic mounted mechanism:
In normal times, the core context is extremely clean, with only the most basic mission objectives; the system will only dynamically mount the corresponding security-audit.skill when Agent's decision “is needed to access the security loophole”; once the audit is completed, the associated context will be unloaded immediately.
Load it, run out. Not only did this save up to 70 percent of the expensive Token for a long-range mission, but it also ensured that the core chain of reasoning remained in high-confidence noise ratio.
Sediment of the bastion into “coded assets”
Prompt is a probate for the model, while Skill is a structured engineering asset (standardized list, trigger conditions, hook test, supporting script). This means that teams can do a Git review, Diff matching, continuous integration and greyscale publish for each Agent Skill, like managing production codes. The certainty of the software project has finally taken over AI uncertainty.
III. The echo of history: ancient engineering discipline became the lifeline of the LLM era
Computer science has developed for more than half a century, and the classic rules that have been drawn up by former human servants — test-driven, single-duty, state-segregated, defensive programming — were first invented not to please computers, but to guard against human programmers' own omissions and confusion.
It is extremely dramatic that today, 50 years later, when we are faced with a LLM that is thousands of times faster in calculation, but is more susceptible to hallucinations, the weapon that can tame it remains this ancient engineering intelligence.
LLM doesn't need you to use all your words in Prompt to teach it how Python wrote it, which already deciphers the world's open source code.
What it lacks is not IQ, but discipline.
From "Hand-written hundreds of thousands of words" Prompt, to "Building Engineering Streaming Lines with Standardized Agent Skills", the current that is going on in GitHub is not just an iterative way of developing tools, but also an AI programming process that is really moving towards a mature bar mitzvah.
♪ Focus on the clean wind of 82, discovering the bottom of technology and ecological evolution ♪
Welcome to the comment area to discuss exchanges and the transmission of ideas.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.