For a long time, I was a happy Cursor user. It felt like a natural extension of VS Code, which I have used for nearly a decade. The completion was fast and precise, the integration was smooth, and it fit well into my existing engineering workflow. Around the middle of last year, I got access through an enterprise license, so I cancelled my personal subscription.
One thing I want to share is that over the past few months, I noticed something interesting: I had quietly stopped using Cursor and moved back to plain VS Code. It was not a deliberate decision at first. I did not sit down and decide that Cursor was no longer useful. I simply found myself opening VS Code more often, and Cursor less often, until the habit had fully shifted.
Part of the reason was that completion became too aggressive for me. In a coding environment, the editor is not just a place where text appears. It is also where thoughts are formed, checked, revised, and sometimes abandoned. When completion interrupts too often, it does not merely add suggestions. It changes the rhythm of thinking. At some point, the assistance started to feel less like support and more like interference.
Another reason was that generating more code inside the editor did not always make me faster. In many cases, it moved the bottleneck from writing code to reviewing code. The scarce resource was no longer typing speed, but attention, trust, and verification. A tool that produces a lot of code also produces a lot of responsibility for the person who has to understand, judge, and maintain it.
The shift in Cursor 3 toward a more chat-centered experience also changed how I evaluated the product. Once the main interaction moves away from the editor and into chat, I naturally start comparing it with Claude, Copilot, and other coding-agent workflows. At that point, the question is no longer only whether the editor experience is better. It becomes whether the new interaction model is strong enough to justify leaving the old one behind.
This illustrates how fragile user trust can be. When a tool sits inside the environment where builders think, write, and review code, small changes matter a lot. If it suggests too much, hides too much, or changes the workflow faster than the value becomes clear, users may not complain loudly. They may simply return to the tools where they feel more in control.
That is the part I find strategically interesting. This is not about whether Cursor is good or bad, but rather how AI tools can evolve from completion to chat to agents without losing the trust and rhythm that made builders adopt them in the first place and continue to use them.
Today, I opened my personal Cursor account, clicked “Upgrade to Pro”, checked the price, and closed the tab. See you next time, Bro.
The following content is generated by LLMs and may contain inaccuracies.
Cursor Adoption Loss Through Workflow Disruption
Context
This note sits at the intersection of developer tooling UX, cognitive psychology of flow states, and product strategy for AI-native tools. It addresses a tension that is becoming structurally significant in 2025–2026: AI coding tools are growing in adoption at a remarkable rate, yet the relationship builders have with those tools is quietly degrading in quality.
Developer favorability toward AI coding tools dropped from over 70% in 2023–2024 to 60% in 2025, even as adoption rates rose to 91%. Developers are using these tools more but trusting them less. The author’s experience — a gradual, unannounced drift back to plain VS Code — is not an edge case. It is a signal that maps onto a broader structural pattern: adoption curves and satisfaction curves are diverging.
The specific mechanism the author identifies is workflow rhythm disruption: the editor is not merely a text-entry surface but a cognitive space where code is thought through, not just written. When AI completion interrupts that rhythm too aggressively, it doesn’t just add noise — it changes the character of the work itself. The second layer — the shift in Cursor 3 toward a chat-centered experience — then forces a product comparison reframe that Cursor may not win on neutral ground.
Key Insights
1. The flow-state disruption problem is empirically documented, not just felt
The author describes how completion that “interrupts too often” changes “the rhythm of thinking.” This matches what the research literature now formally measures. Mental flow is a well-established psychological construct defined as a state of energized focus and full involvement, and is a core determinant of developer productivity in both academic and industrial frameworks. Empirical studies consistently show that maintaining uninterrupted flow yields substantial productivity gains, while even brief interruptions incur disproportionate recovery costs.
More specifically, a 2025 study of real-world commits found that 68.81% of model recommendations disrupt developers' ongoing mental flow, including 8.83% of suggestions that are technically correct but ill-timed — confirming the author’s intuition that the problem isn’t just quality of suggestions, but timing. A correct suggestion at the wrong moment is still a disruption.
Research on completion acceptance patterns corroborates this: typing speed and the presence or absence of pauses provide insight into the developer’s cognitive state. Sustained high-speed typing with minimal pauses suggests focus or flow — a state in which the developer is less likely to welcome external suggestions. In contrast, slower or fragmented typing often coincided with a higher likelihood of suggestion acceptance.
2. The attention-as-bottleneck insight is backed by verification-load research
The author makes a precise claim: generating more code moved the bottleneck from typing to reviewing — “the scarce resource was no longer typing speed, but attention, trust, and verification.” A 2026 CHI paper formalizes this as “verification load.” This operationalizes extraneous cognitive load and flow disruption in a form that travels across interaction styles and backends. With the same backend, interface alone materially shifts the assistance–burden trade-off. The cost of checking and repairing model output is a distinct cognitive tax that accumulates across repeated use and produces stress and fatigue — not visible in lines-of-code metrics.
3. The METR RCT: the productivity perception gap
A METR randomized controlled trial conducted in July 2025 measured 16 experienced open-source developers completing 246 real-world issues across massive repositories. The data revealed that developers using AI tools were 19% slower than developers working without AI assistance. A significant perception gap emerged: participants believed AI tools made the coding process 20% faster, creating a 40 percentage point difference between perceived and actual performance. This matters for the author’s narrative: silent drift back to VS Code may be the body’s honest accounting, even when the mind still expects AI to help.
4. The trust–adoption divergence is structural, not individual
Developer trust in AI is declining even as adoption rises. In 2023 and 2024, more than 70% of developers expressed positive sentiment toward AI tools. By 2025, that number dropped to 60%. Only 33% trust AI-generated code for accuracy. 46% actively distrust it. This describes a population engaged in something they don’t fully trust: 84% use the tools or plan to, while a third say they don’t believe the output. This is not the profile of a satisfied customer base. It’s the profile of a workforce that feels it has no choice.
5. Cursor’s strategic pivot to chat-then-agents changed the comparison set
The author astutely notices that once the main interaction surface moved from the editor to chat, the comparison shifted from editor quality to agent quality — and Cursor no longer had a home-field advantage. In March 2025, users of Cursor’s Tab autocomplete outnumbered agent users 2.5 to 1. That ratio has now reversed: agent users outnumber Tab users 2 to 1. “Cursor is no longer primarily about writing code,” according to Cursor’s own leadership.
Once evaluated as an agent, Cursor competes on different terrain. Cursor doesn’t outperform any competitor on any single dimension. On planning, Claude Code is stronger. On autonomous reasoning, Codex is stronger. On code generation alone, the four tools are about the same. The author’s instinct — that moving to chat forces a re-evaluation — reflects the actual competitive reality.
6. Claude Code and Codex as the natural alternatives once chat becomes primary
Claude Code is Anthropic’s command-line coding tool. It runs in a terminal alongside a developer’s normal workspace and connects to Claude’s models, with a 1M-token context window. That means it can hold most of a codebase in memory at once. Of the four major tools, Claude Code has the strongest contextual awareness across an entire codebase.
A pragmatic pattern is already emerging in enterprise: heavy lifting — large refactors, writing test suites across dozens of files, CI/CD automation — goes to Claude Code; interactive editing and day-to-day file editing, quick bug fixes, UI work, and reviewing code goes to Cursor. Tab completions make line-by-line editing fast. The author’s personal story may be resolving into exactly this dual-tool equilibrium — VS Code (or Cursor’s core) for thinking-in-code, an agent for delegated tasks.
7. The pricing controversy as an additional trust-eroding event
The author’s final scene — checking the Pro upgrade price and closing the tab — is not trivial. It occurs in a specific historical moment when Cursor’s pricing changes had already burned trust with power users. In June 2025, Cursor introduced changes to how the Pro plan worked. Users reported logging in to find their plan had effectively changed without clear advance notice, or that the new terms were buried in documentation. The new structure meant that some workflows that had been comfortably within the Pro plan limits suddenly weren’t. Heavy users reported $10–20 daily overages. One team’s $7,000 annual subscription depleted in a single day. The economic uncertainty compounds the cognitive one.
8. The enterprise lock-in paradox
The author’s usage pattern — enterprise license removes the personal subscription incentive — reflects a broader dynamic. The company’s revenue mix moved from consumer/individual seats toward enterprise contracts over 2025. Corporate buyers grew from ~25% of revenue in late 2024 to ~45% at $1B ARR and toward ~60% at $2B ARR. Enterprise licenses can paradoxically reduce personal investment: when an individual cancels their personal subscription after getting access through work, they lose the skin-in-the-game that drives deeper adoption. They become passive users, more susceptible to drift.
9. The “Cursor as identity” advantage is fragile for expert users
Cursor’s product-led growth was built on a specific user type: the strategy was to serve the “10x user” — not the average user, but the most demanding user in the category. The user who will restructure their workflow around a product if it is good enough. These users pay more, evangelize more, and are harder to displace. But the author represents exactly this profile — a decade-long VS Code user who adopted early and deeply — and they are precisely the ones most sensitive to rhythm disruption. The more expert the user, the lower the tolerance for unsolicited interference.
Open Questions
1. Is “invisible churn” a dark pattern in AI tool metrics? Aggregate DAU and ARR look healthy for Cursor, but the author’s experience — enterprise-covered, not officially churned, yet effectively no longer using the product — may represent a class of users that standard retention metrics cannot see. How much of Cursor’s enterprise ARR is held by organizations whose engineers have silently reverted to old habits? Could the real adoption signal be the ratio of active AI-assisted PRs per seat, rather than seat count?
2. Can an AI coding tool be designed to read the developer’s cognitive state and withdraw suggestions — not just offer them? The research on typing rhythm suggests that developers telegraph their flow state through behavioral signals. The EditFlow benchmark shows that even technically correct suggestions disrupt flow 68.81% of the time. Is there a design space between “always-on completion” and “chat-on-demand” that adjusts suggestion aggressiveness in real time based on detected cognitive load — and would developers actually want a tool that does less on purpose?
很长一段时间里，我是一个快乐的 Cursor 用户。它感觉像是 VS Code 的自然延伸，而我已经使用 VS Code 近十年了。代码补全快速精准，集成流畅，完全融入了我现有的工程工作流。去年年中左右，我通过企业许可证获得了访问权限，所以取消了个人订阅。
我想分享的一件事是，在过去的几个月里，我注意到了一些有趣的现象：我悄悄地停止了使用 Cursor，转而回到了普通的 VS Code。这不是一个深思熟虑的决定。我没有坐下来决定 Cursor 不再有用。我只是发现自己越来越经常地打开 VS Code，越来越少地打开 Cursor，直到这个习惯完全改变了。
原因之一是代码补全对我来说变得太积极了。在编程环境中，编辑器不仅仅是文本出现的地方。它也是思想形成、检查、修改，有时被放弃的地方。当补全太频繁地打断时，它不仅仅是添加建议。它改变了思考的节奏。在某个时刻，这种辅助开始感觉不像是支持，而更像是干扰。
另一个原因是在编辑器内生成更多代码并不总是让我工作得更快。在很多情况下，它将瓶颈从代码编写转移到了代码审查。稀缺的资源不再是打字速度，而是注意力、信任和验证。一个产生大量代码的工具也会产生大量责任，需要使用者去理解、判断和维护这些代码。
Cursor 3 向以聊天为中心的体验的转变也改变了我对产品的评价方式。一旦主要交互从编辑器转向聊天界面，我自然会开始将它与 Claude、Copilot 和其他代码代理工作流进行比较。此时，问题不再仅仅是编辑器体验是否更好。它变成了新的交互模式是否足够强大，足以证明离开旧方式的合理性。
这说明了用户信任有多脆弱。当一个工具存在于建筑师思考、编写和审查代码的环境中时，小的改变意义重大。如果它建议过多、隐藏过多，或改变工作流的速度快于价值显现的速度，用户可能不会大声抱怨。他们可能只是简单地回到那些让他们感觉更能掌控的工具。
这正是我认为在战略上有趣的地方。这不是关于 Cursor 好不好的问题，而是关于 AI 工具如何能够从代码补全演进到聊天，再到代理，同时不失去最初驱动建筑师采纳它并继续使用它的信任和节奏。
今天，我打开了我的个人 Cursor 账户，点击了"升级到 Pro"，查看了价格，然后关闭了标签页。下次见，伙计。
以下内容由 LLM 生成，可能包含不准确之处。
光标工具采用流失与工作流中断
背景
本笔记位于开发者工具 UX、心流状态的认知心理学和AI 原生工具的产品战略的交汇处。它解决了在 2025–2026 年间变得结构性显著的一个张力：AI 编码工具采用率在以惊人的速度增长，但开发者与这些工具的关系质量却在悄然恶化。
开发者对 AI 编码工具的好感度从 2023–2024 年的 70% 以上下降到 2025 年的 60%，尽管采用率上升到 91%。开发者在更多地使用这些工具，但信任度却在下降。作者的经验——逐渐、无声地漂移回纯 VS Code——不是边界情况。它是一个映射到更广泛结构模式的信号：采用曲线和满意度曲线正在背离。
作者识别的具体机制是工作流节奏中断：编辑器不仅仅是文本输入表面，而是一个认知空间，在这里代码是被思考的，而不仅仅是被写出来的。当 AI 完成建议过于激进地中断这种节奏时，它不仅仅是增加噪声——它改变了工作本身的性质。第二层——Cursor 3 向聊天中心体验的转变——随后强制了一个产品比较的重新框架，Cursor 可能无法在中立立场上赢得这场比较。
关键洞察
1. 心流状态中断问题有实证记录，不仅仅是主观感受
作者描述了完成建议"过于频繁地中断"如何改变了"思考的节奏"。这与研究文献现在正式衡量的内容相符。心流是一个已建立的心理学概念，定义为精力充沛的专注和充分投入的状态，也是开发者生产力的核心决定因素，在学术和工业框架中都是如此。实证研究一致表明，保持不间断的心流会产生实质性的生产力收益，而即使是简短的中断也会产生不成比例的恢复成本。
更具体地说，2025 年对真实提交的研究发现，68.81% 的模型建议会中断开发者的持续心流，其中 8.83% 的建议在技术上是正确的，但时机不当——证实了作者的直觉，即问题不仅仅是建议的质量，还有时机。一个在错误时刻的正确建议仍然是一个中断。
关于完成接受模式的研究验证了这一点：打字速度以及是否存在暂停为开发者的认知状态提供了洞察。持续的高速打字伴随最少的暂停表明专注或心流——在这种状态下，开发者不太可能欢迎外部建议。相比之下，较慢或零碎的打字往往与更高的建议接受可能性一致。
2. 注意力作为瓶颈的洞察得到验证负荷研究的支持
作者提出了一个精确的论点：生成更多代码将瓶颈从打字转移到了审查——“稀缺资源不再是打字速度，而是注意力、信任和验证”。一篇 2026 年 CHI 论文将其形式化为"验证负荷"。这在跨交互风格和后端的形式中体现了额外的认知负荷和心流中断。使用相同的后端，仅界面就实质性地改变了辅助-负担权衡。检查和修复模型输出的成本是一种不同的认知税，在重复使用中积累，并产生压力和疲劳——在代码行指标中看不见。
3. METR 随机对照试验：生产力认知差距
METR 在 2025 年 7 月进行的随机对照试验测量了 16 名经验丰富的开源开发者完成 246 个跨大型存储库的真实问题。数据显示，使用 AI 工具的开发者比不使用 AI 辅助的开发者慢 19%。出现了显著的认知差距：参与者认为 AI 工具使编码过程快 20%，造成了 40 个百分点的认知与实际性能差异。这对作者的叙述很重要：无声漂移回 VS Code 可能是身体的诚实记账，即使心智仍然期待 AI 能有帮助。
4. 信任–采用背离是结构性的，不是个体的
开发者对 AI 的信任正在下降，即使采用率在上升。在 2023 和 2024 年，超过 70% 的开发者表达了对 AI 工具的积极情绪。到 2025 年，这个数字下降到 60%。只有 33% 的人信任 AI 生成代码的准确性。46% 积极不信任它。这描述了一个从事他们不完全信任的工作的人口：84% 使用这些工具或计划使用，而三分之一的人说他们不相信输出。这不是满意客户群的形象。这是一个感到没有选择的劳动力的形象。
5. Cursor 向聊天-代理体验的战略转向改变了比较集合
作者敏锐地注意到，一旦主要交互表面从编辑器转移到聊天，比较就从编辑器质量转变为代理质量——Cursor 不再有主场优势。在 2025 年 3 月，Cursor 标签自动完成的用户数量超过代理用户 2.5 倍。该比例现已反转：代理用户超过标签用户 2 倍。根据 Cursor 自身领导力的说法，“Cursor 不再主要是关于编写代码”。
一旦被评估为代理，Cursor 就在不同的地形上竞争。Cursor 在任何单一维度上都不优于任何竞争对手。在规划方面，Claude Code 更强。在自主推理方面，Codex 更强。在代码生成本身上，四个工具大致相同。作者的直觉——聊天的转移强制重新评估——反映了实际的竞争现实。
6. 一旦聊天成为主要形式，Claude Code 和 Codex 作为自然的替代方案
Claude Code 是 Anthropic 的命令行编码工具。它在终端中与开发者的常规工作区域并行运行，并连接到 Claude 的模型，具有 100 万令牌的上下文窗口。这意味着它可以一次在内存中保存大多数代码库。在四个主要工具中，Claude Code 对整个代码库具有最强的上下文意识。
企业中已经出现了一个务实的模式：繁重工作——大型重构、跨数十个文件编写测试套件、CI/CD 自动化——转到 Claude Code；交互式编辑和日常文件编辑、快速错误修复、UI 工作和代码审查转到 Cursor。标签完成使逐行编辑快速。作者的个人故事可能正在解决为完全这样的双工具平衡——VS Code（或 Cursor 的核心）用于编码中的思考，一个代理用于委托任务。
7. 定价争议作为额外的信任侵蚀事件
作者的最后一幕——检查专业升级价格并关闭标签——并非微不足道。它发生在一个特定的历史时刻，当时 Cursor 的定价变化已经用力量用户的信任。2025 年 6 月，Cursor 推出了对专业计划工作方式的更改。用户报告登录时发现他们的计划在没有明确提前通知的情况下实际上已更改，或者新条款被埋在文档中。新结构意味着一些曾经舒适地在专业计划限制内的工作流突然不是了。重度用户报告每日超支 10-20 美元。一个团队的 700 万美元年度订阅在一天内耗尽。经济不确定性加剧了认知上的不确定性。
8. 企业锁定悖论
作者的使用模式——企业许可证消除了个人订阅的激励——反映了更广泛的动态。公司的收入组合在 2025 年从消费者/个人座位转向企业合同。企业客户从 2024 年末的约 25% 收益增长到 10 亿美元 ARR 时的约 45%，并朝着 20 亿美元 ARR 时的约 60% 发展。企业许可证可能会自相矛盾地减少个人投资：当个人通过工作获得访问权限后取消他们的个人订阅时，他们失去了推动更深层采用的皮肤利益。他们成为被动用户，更容易漂移。
9. “Cursor 作为身份"优势对专家用户是脆弱的
Cursor 的产品主导增长是为特定用户类型而建立的：战略是服务"10 倍用户”——不是平均用户，而是该类别中最苛刻的用户。如果足够好，将重组他们的工作流以适应产品的用户。这些用户支付更多，进行更多倡导，更难被替换。但作者代表了完全这个档案——十年的 VS Code 用户，早期并深度采用——他们正是对节奏中断最敏感的人。用户越专业，对无故干扰的容忍度就越低。
开放问题
1. “隐形流失"是 AI 工具指标中的暗模式吗？
总体 DAU 和 ARR 对 Cursor 来说看起来健康，但作者的经验——企业覆盖，未正式流失，但有效地不再使用该产品——可能代表一类标准保留指标无法看到的用户。Cursor 的企业 ARR 中有多少是由工程师无声地恢复到旧习惯的组织持有的？真正的采用信号可能是每座位的活跃 AI 辅助 PR 比率，而不是座位数？
2. AI 编码工具能否被设计成读取开发者的认知状态并撤回建议——而不仅仅是提供建议？
关于打字节奏的研究表明，开发者通过行为信号透露他们的心流状态。EditFlow 基准表明，即使在技术上正确的建议也会 68.81% 的时间中断心流。在"始终打开的完成"和"按需聊天"之间是否存在一个设计空间，可以根据检测到的认知负荷实时调整建议的激进性——开发者是否真的想要一个故意做少的工具？