摘要：2026年，Prompt一边被高管宣判“已死”，一边在闲鱼上被拆成几块钱的模板、几千块的接单服务卖掉。猎聘口径下Prompt岗位同比增长486.84%，平均年薪34.31万元；但另一面，Prompt工程师岗位数量从峰值下降67%，平均年薪两年内腰斩。两组数字并不矛盾：“背神级咒语”的套利空间在被吃掉，而“把模糊需求翻译成可执行规格、放进可验证循环”的能力正在扩散成基础工作能力。
一、Prompt的六年：从“给例子”到“编译程序”
CoT让模型“会想”，ReAct让模型“能查”，RAG把信息环境本身变成Prompt的一部分——Prompt从“指令措辞”扩展到了“上下文设计”。
二、自动优化：DSPy与GEPA
DSPy（ICLR 2024 Spotlight）把APE/OPRO的“搜一条指令”扩展为“编译一个程序”：定义模块、指标和训练数据，编译器搜索指令和示例的组合。截至2026年中期，DSPy约34,000+ GitHub stars，Databricks、VMware、JetBlue在生产中使用。文档记录的案例：ReAct准确率从24%提升到51%，RAG流水线从53%提升到61%。
GEPA（ICLR 2026 Oral）走另一条路：用自然语言反思从执行轨迹中学习，而不是依赖标量奖励。部分任务上比GRPO平均高10%、最高20%，同时使用35倍更少的rollout。一个反直觉的边界：20-100个样本的“轻量模式”反而优于500个样本，说明GEPA的优势在数据稀缺、指标清晰的场景中更佳。
三、Agent循环与Prompt Injection：Prompt变成了程序
2024年之后，Prompt不再只是“对模型说的一段话”，而是定义“推理-行动-观察”循环、工具使用规则、记忆和检索结构的程序。Claude Code之父Boris Cherny的总结被反复引用：Prompt脚手架是一种“会过期的负债”。Opus 5发布时，他的团队用消融法砍掉了Claude Code超过80%的系统提示词，模型表现几乎没有变化——删掉的不是“把任务说清楚”，而是“为上一代模型打的补丁”。
另一面是安全。OWASP LLM Top 10在2025年将Prompt Injection列为LLM01，即LLM应用的首要安全风险。它的机制不是“模型不够聪明”，而是指令和数据共享同一个文本通道，任何被读取的文本都可能变成指令，没有语法层面的修复方案。防御需要系统级设计：指令层级、输入隔离、最小权限、信任边界管理。
四、2026年的能力转向
“会写Prompt”作为独立岗位的套利空间在收窄，但“把模糊需求翻译成可执行规格”的能力正在扩散成一种基础工作能力——它嵌在产品、运营、设计、创业的每一个环节里，只是不再单独标价。
闲鱼上的AI轻创业数据印证了另一条路径：AI技能接单占全部AI订单的45.1%，AI教程与课程占8.1%，AI模板与工作流占6.6%。这些路径的共同点不是“会写神级提示词”，而是把AI能力封装成一个可交付的产品或服务。
博主走通并验证的一条链路是：Langfuse管版本和追踪 → Promptfoo做回归和错误监控 → 有干净指标后接GEPA做自动优化。2026年关于Prompt结构最稳定的共识，是把一条Prompt拆成五个可检查的要素：角色、任务、上下文、约束、输出格式。
查看原文
以上仅为核心速览。完整版包含：2020-2028年Prompt方法论完整时间线、CoT/Self-Consistency/ToT/ReAct/RAG的逐项拆解与数据、DSPy与GEPA的完整对比、Prompt Injection的系统级防御框架、猎聘与闲鱼两组数据的完整分析，以及2026年Prompt技术最实用的开源资源汇总。
本文首发于微信公众号「阿Q的AI笔记」，点击下方链接阅读全文：
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。
summary: In 2026, Prompt was sentenced by the executive to “dead” while he was sold a few dollar-scraping templates and thousands of dollars of billing services on idle fish. On the other hand, the number of Prompt engineers has fallen from a peak of 67 per cent and their average annual salary has been cut off within two years. The two sets of figures are not contradictory: the arbitrage of the “back-level spell” is being eaten, and the ability to “translate vague demand into enforceable specifications and put it into a verifiable cycle” is spreading into basic working capability.
Prompt's six years: from “to give examples” to “to compile”
COT allows the model “to think”, ReAct allows the model “to look”, and RAG expands the information environment itself into a part of Prompt — Prompt from “directive wording” to “context design”.
Automatic optimization: DSPy and GEPA
DSPy (ICLR 2024 Spotlight) extended the APE/OPRO “search a directive” to “compile a program”: define modules, indicators and training data, compiler search instructions and sample combinations. By mid-2026, about 34,000+ GitHub Stars, Databricks, VMware and JetBrue were used in production. Documented cases: React's accuracy increased from 24 per cent to 51 per cent, and RAG's current line from 53 per cent to 61 per cent.
GEPA (ICLR 2026 Oral) takes a different path: to reflect in natural languages on learning from the implementation trajectory, rather than relying on a target incentive. Some of the tasks are 10 per cent higher and up to 20 per cent higher on average than GRPO, while 35 times less rollout is used. An anti-intuitive border: the “light mode” of 20-100 Z111Z is better than 500 Z111Z, suggesting that the GEPA advantage is better in the context of scarce data and clear indicators.
III. Agent Looping and Prompt Injection:Prompt became a program
After 2024, Prompt is no longer merely a “say to the model”, but rather a procedure for defining the “debate-action-observation” cycle, the rules for the use of tools, memory and retrieval structures. The summary of Boris Cherny, the father of Claude Code, was quoted repeatedly as saying that the Prompt scaffold was an “expired liability”. When Opus 5publish was used by his team to cut down more than 80 per cent of Claude Code's system, prompt, the model's performance was virtually unchanged — not “defining the mission”, but “compensating the previous generation's model”.
The other side is security. OWASP LLM Top 10 classified Prompt Injection as the primary security risk for LLM01 applications in 2025. Its mechanism is not that “models are not smart enough”, but the same text channel for command and data sharing, and any text that is read may become a directive and there is no grammatical repair programme. Defense requires system-level design: command level, input segregation, minimum permission, confidence border management.
IV. Capacity shift in 2026
The "prompt" arbitrage space as a stand-alone position is being narrowed, but the ability to "translate vague demand into enforceable specifications" is spreading into a basic work capability — embedded in every link of product, operation, design, entrepreneurship, but no longer priced separately.
The AI light entrepreneurship data on idle fish confirm another path: AI skill lists account for 45.1 per cent of all AI orders, AI curricula and courses 8.1 per cent, AI templates and workflow 6.6 per cent. The common denominator of these paths is not “the writing of God-class prompt”, but the encapsulation of AI capabilities as a deliverable product or service.
One of the links through and validated by the blogger is the Langfuse version and tracking the Promptfoo return and error monitoring, and then the GEPA automatically optimized with clean indicators. The most stable consensus on the Prompt structure in 2026 was the dismantling of a Prompt into five searchable elements: role, mission, context, binding, output format.
View Original
The above is only a core snapshot. The complete version consists of the complete time line of the Prompt methodology 2020-2028, the itemized dismantling and data of CoT/Self-Consistency/ToT/React/RAG, the complete comparison of DSPy and GEO, the system-level defence framework of Prompt Injection, the complete analysis of the two sets of data on hunting and idle fish, and the most practical summary of open source resources of Prompt technology 2026.
This post was first sent to WeChatWeChat Official Account, "AQ's AI Notes", and is read in full by clicking on the following link:
Auto-fetched and translated by AI Atlas. All rights belong to the original author.