June 6, 2026

AI Engineer founder and Latent Space host Shawn Wang (swyx) offers a sharper alternative to always running coding agents in "plan mode": frame your task as a question rather than a command. Phrasing it as a question invites the model to push back, rate the quality of the idea, and suggest alternatives instead of blindly executing what you literally said, which is often not what you meant. He notes that literally just appending a "?" to the end of your prompt often does the trick.

AI Engineer 创始人、Latent Space 主播 Shawn Wang(swyx)给出了一个比"永远开 plan 模式"更聪明的做法:把任务写成一个问句,而不是一条命令。用提问的方式表达,会引导模型反驳你、评估这个想法的好坏、提出替代方案,而不是机械地照你字面说的去做,而你字面说的往往并不是你真正想要的。他说,有时仅仅在 prompt 末尾加一个"?"就足够了。

Anthropic's Boris Cherny, who works on Claude Code, says Claude Cowork is at its best on work that's too big for a chat window: research across dozens of accounts, recurring reports, triaging an inbox and drafting replies. Anthropic has doubled Cowork usage limits for the next month against the 5-hour rate limits, so he frames this as a good month to hand off a big, messy project you've been saving up.

Anthropic 负责 Claude Code 的 Boris Cherny 说,Claude Cowork 最擅长那些"塞不进一个聊天窗口"的活儿:跨几十个账户做调研、生成周期性报告、整理收件箱并起草回复。Anthropic 在接下来一个月把 Cowork 的用量上限翻了一倍(针对 5 小时的速率限制),所以他说这正是一个把你攒了很久的大型、杂乱项目交出去的好时机。

OpenAI's Thibault Sottiaux, who works on Codex and ChatGPT, is highlighting two threads of progress: Codex "papercuts" going down while adoption goes up, and a memory improvement he sums up as "Better memory = Shorter prompts = More utility per token." The throughline is squeezing more usefulness out of each token by reducing the context users have to re-supply.

OpenAI 负责 Codex 和 ChatGPT 的 Thibault Sottiaux 强调了两条进展:Codex 的"小毛刺"在减少、采用率在上升;以及一项他总结为"更好的记忆 = 更短的 prompt = 每个 token 更高效用"的记忆改进。核心思路是通过减少用户需要反复补充的上下文,让每个 token 发挥更大价值。

Roblox product leader Peter Yang shared a five-step recipe for building AI skills that check their own work and improve over time: give the skill context with examples of good output, write a clear "Use when..." trigger description, add an evals file with about ten pass/fail checks, add a memory file that captures one-sentence learnings from past chats, and finally build a meta-skill that cleans up other skills by stripping duplicate or stale instructions and AI slop. He also interviewed builder Mike Van Horn, who has no CS degree yet ships projects and has contributed to repos like Python and Go.

Roblox 产品负责人 Peter Yang 分享了一套五步法,用来构建"会自我检查、并随时间改进"的 AI skill:用好输出的示例给 skill 提供上下文;写一个清晰的"Use when..."触发描述;加一个含约十条通过/失败检查的 evals 文件;加一个 memory 文件,用一句话记录过去对话中的经验;最后再做一个元 skill,专门清理其他 skill,去掉重复或过时的指令以及 AI 废话。他还采访了开发者 Mike Van Horn——此人没有 CS 学位,却持续交付项目,并为 Python、Go 等仓库做过贡献。

Former Google Gemini and Veo product leader Madhu Guru has a contrarian warning for enterprise AI teams: stop building for today's model capabilities and price points. Instead, think six months out, assume models will be far smarter and cheaper, and scaffold around today's weaknesses to push the frontier, betting the next generation natively solves whatever the scaffold patches. "Build for the slope," he says, arguing that the repeated ability to identify and bridge model gaps becomes a moat of its own.

前 Google Gemini、Veo 产品负责人 Madhu Guru 给企业 AI 团队提了个反共识的警告:别再围绕"今天的模型能力和价格"去构建。要往六个月后想,假设模型会变得更聪明也更便宜,围绕今天的短板搭脚手架去推进前沿,并赌下一代模型会原生地解决掉脚手架所修补的问题。他说要"为这条上升曲线而建",并认为这种反复识别并弥合模型差距的能力本身就会成为护城河。

Vercel CEO Guillermo Rauch announced two infrastructure pieces. First, a novel virtual storage layer where agent filesystem state can be read, written, and mounted independently of a Sandbox's lifecycle — storage that is decoupled but attachable to Builds, Functions, Sandboxes and more. Second, a Skills API he describes as "the npm registry for agent capabilities and extensibility," free and open, meant to make any agent or platform smarter.

Vercel CEO Guillermo Rauch 公布了两块基础设施。其一是一套全新的虚拟存储层:agent 的文件系统状态可以独立于 Sandbox 生命周期被读取、写入和挂载——存储与计算解耦,但可挂接到 Builds、Functions、Sandboxes 等之上。其二是一个 Skills API,他形容为"agent 能力与可扩展性的 npm registry",免费且开放,目的是让任何 agent 或平台都变得更聪明。

Box CEO Aaron Levie argues that coding is basically the pinnacle of what AI can automate today, and yet human engineers are still needed to oversee the agents. He lists why coding is uniquely favorable: models are trained on enormous amounts of sophisticated code, users are highly technical, the work is verifiable because you can test an app, outcomes are forgiving (sloppy code can still ship a working app), and the context already sits digitized in the codebase. Most knowledge work lacks those advantages, so if engineers remain in high demand even here, he concludes the displacement risk for other knowledge work is lower than perceived — agents let people do far more, but the people don't go away.

Box CEO Aaron Levie 认为,写代码基本上是当今 AI 能自动化的"天花板",但即便如此,仍然需要人类工程师来监督 agent 才能让它们真正有效。他列出了写代码为何格外占优:模型在海量复杂代码上训练过、用户本身就很技术、工作可验证(因为你能测试一个 app)、结果容错(代码很糙也能跑出能用的 app)、而且上下文早已数字化地存在于代码库里。大多数知识工作并不具备这些优势,所以如果连写代码这种最有利的场景都仍然高度需要工程师,他的结论是:其他知识工作被取代的风险其实比外界以为的要低——agent 会让人做得更多,但人不会消失。

Cursor design lead Ryo Lu showed off direct manipulation for designing in code: click, chat, and hold shift to multi-select. He says it works best with Composer 2.5 — a step toward making in-code design feel as fluid as moving objects on a canvas.

Cursor 设计负责人 Ryo Lu 展示了在代码里做设计的直接操作方式:点击、对话、按住 shift 多选。他说配合 Composer 2.5 体验最佳——这是让"在代码中做设计"变得像在画布上拖拽对象一样顺手的一步。

South Park Commons general partner Aditya Agarwal offered a one-line gut check for builders: "Sometimes speed is just impatience disguised as ambition." A reminder that moving fast is not always the same thing as moving toward the right thing.

South Park Commons 合伙人 Aditya Agarwal 给开发者抛出一句扎心的自检:"有时候,所谓的快只是把不耐烦伪装成了野心。"这是在提醒:跑得快,未必等于朝着对的方向跑。

Anthropic's Claude account announced that Claude Cowork usage limits are doubled for the next month, live on all paid plans through July 5, encouraging users to delegate bigger and more complex tasks. The download is via the Claude desktop app.

Anthropic 的 Claude 官方账号宣布,Claude Cowork 的用量上限在接下来一个月翻倍,所有付费方案均可用,持续到 7 月 5 日,鼓励用户把更大、更复杂的任务交给 Claude。可通过 Claude 桌面端 app 下载使用。

The Takeaway: The so-called "SaaS apocalypse" is backwards — as AI turns nearly everyone into a builder, the hard part stops being writing code and becomes running software reliably, which makes established SaaS a gold mine, not a graveyard.

Matt Colyer, Figma's director of product management for developers, has been building personal AI agents for two years, long before "vibe coding everything" became the dominant January narrative. His through-line: it is genuinely fun to build the first version of a tool, but software companies sell far more than code. He now buys more software than ever, because once you live with the ongoing cost of maintaining your own rickety agent, paying someone else to run it starts to look like a bargain. "I'm buying more software these days than I ever did before... I'm just gonna pay somebody else to run the agent for me."

His sharpest reframing is that in the AI era, "every problem becomes a context problem" — the real work is framing a problem with the right information. He describes a Figma internal system that walks the org chart, reads the last thirty days of relevant Slack channels, and checks the Asana board to auto-generate uncannily good onboarding docs for new hires, pulling knowledge that used to live only in a manager's head.

On design specifically, he pushes past the chat box: agents on Figma's infinite canvas can support both divergent thinking (an agent that throws many concept frames at you) and convergent thinking (an agent that clusters and critiques them like a customer would). And he's blunt that the 2026 bottleneck has shifted from generation to review — we now have cheap agents flooding us with content, and the unsolved problem is how to scale our own value system to evaluate it with enough trust to let some of it run in auto mode.

要点:所谓的"SaaS 末日论"其实把因果搞反了——当 AI 让几乎人人都能成为 builder,难点就不再是写代码,而是把软件可靠地运行起来,这恰恰让成熟的 SaaS 成了金矿,而不是坟场。

Matt Colyer 是 Figma 面向开发者的产品管理总监,他做个人 AI agent 已经两年了,远早于今年一月"什么都拿来 vibe code"成为主流叙事之前。他的核心观点是:做一个工具的第一个版本确实很好玩,但软件公司卖的远不只是代码。如今他买的软件比以往任何时候都多,因为一旦你切身承受了维护自己那套摇摇晃晃的 agent 的持续成本,花钱请别人来替你运行就开始显得很划算。"我现在买的软件比以前任何时候都多……我宁愿花钱让别人替我跑这个 agent。"

他最犀利的重新定义是:在 AI 时代,"每一个问题都变成了上下文问题"——真正的工作是用正确的信息把问题框定好。他描述了 Figma 内部的一套系统:它会遍历组织架构图、读取相关 Slack 频道过去三十天的内容、再查 Asana 看板,自动为新员工生成好得惊人的 onboarding 文档,把过去只存在于管理者脑子里的知识提取出来。

具体到设计,他跳出了聊天框:Figma 无限画布上的 agent 既能支持发散思维(一个不断往你面前抛各种概念框的 agent),也能支持收敛思维(一个像客户那样把它们聚类并评判的 agent)。他还直言,2026 年的瓶颈已经从"生成"转移到了"审阅"——如今廉价的 agent 把内容潮水般地推给我们,而尚未解决的难题是:如何把我们自己的价值判断体系规模化,让我们足够信任它,从而放手让一部分内容进入自动模式运行。