Pinned Tweet
来看看你的1024关键词
请安装 github.com/dusthappy/1024 中的 programmer-day-1024 技能到你当前支持的技能目录，确认可用后，为我生成一份带八维雷达图、夸赞关键词和专属口号的 1024 程序员节礼物。
让 AI 直接开写，返工账单可能更贵。
先让深上下文 agent 写规格，再找零上下文的来挑刺。我觉得这步值得花 token：规格里没写清的接口，等代码写完再改更疼。
Check out Google's Elephant-Goldfish Model, our new open-source plugin → goo.gle/480yn2n
Draft specs with deep-context agents and stress-test them with zero-context subagents before writing code. You make the architectural decisions and approve the spec before implementation. Works with Antigravity, Claude, Codex and ChatGPT.
省一次确认，可能要换一遍密钥。
Claude Code 的「Don't Ask Again」会把密码和 API key 明文写到本地。Apiiro 在 1,994 个公开仓库里找到了 2,539 个暴露密钥。
独立开发图省事很容易一路点下去。文件删了，Git 历史还在；泄露后得先换密钥。
‼️ Claude Code's "Don't Ask Again" option can save passwords and API keys in plaintext.
Apiiro found 2,539 exposed secrets across 1,994 public GitHub repos, including database credentials, SSH logins, and cloud keys.
Deleting files doesn't erase Git history.
Read: thehackernews.com/2026/11/cl…
别让用户靠逆向猜额度。
订阅者称 5x、20x 按每次会话计算，每周额度却低得多；相关索赔获准继续审理。
做学习产品，我会把额度的时间窗口写清楚：省下的说明，迟早要用客服和信任来还。
别把没人用的功能，优化到天荒地老。
代码能写完，bug 还在。做学习产品，我更愿意先看用户卡在哪一步、翻一下线上 trace，再决定让 agent 改什么。token 花出去，至少得换回用户在乎的东西。
I unironically agree with @bcherny’s take that “coding is solved, but bugs are not.”
The deeper meaning is that you can in theory code *anything* now. You can burn infinite tokens refining and optimizing. Including optimizations that no one will care about 💀
With Vercel you can build a different kind of recursive improvement loop. You capture traces from production, you then optimize and course-correct based on how your software is being used IRL.
This is not a new idea. The best engineers I’ve ever worked with never jumped into coding (now prompting) right away. You instrument the system, collect data, understand distributions, user journeys, and then put fingers on keyboard.
You can ask your agent today to look at your traces!
省掉一次登录，可能赔上整个已登录会话。
Agent 可以直接操作带着标签页、登录状态和扩展的浏览器。做独立产品，我会先想清楚哪些页面能交给它。
误操作比省登录更贵。
OpenCode Browse is now available
it lets your agent control your real browser with your tabs and logins and extensions
opencode.ai/product/browse
别先算省了多少代码，依赖能不能继续用更要命。
Asahi 把封闭的平台打开，新增限制却让其他项目更难接着做；每月 $100 的赞助也被取消了。
做独立产品，依赖的边界得看清。
别把能跑的演示当成能维护的产品。
一条提示词生成 Three.js 场景，塞进一个 HTML，验证想法确实快。独立开发真要接着改交互、查性能，代码全挤在一起的代价很快就来了。
GPT-6.2 Sol is already showing up for some ChatGPT users, and just look at what it's creating.
These aren't screenshots from a game. They're Three.js scenes generated from a single prompt, with everything packed into ONE HTML file.
I've been waiting to see what Sol could do. And now I think we're finally going to see a worthy competitor to Opus 5.5 for Plus users.
最贵的功能，是别人已经做完的那一套。
t3os 还没收尾，Omarchy 已经覆盖了它想做的大部分 agent 能力。独立开发先拿现成工具跑一遍，能省下的不是几行代码，是后面整套维护成本。
别先让 Agent 写代码。
注册就可能卡一下午。
1,035 次测试只有 16.6% 跑出第一个可用结果，验证码拦了 373 次。独立开发接数据库时，得把注册和拿凭据的人工步骤也算进时间。
How much revenue are database companies losing because agents aren’t able to sign up autonomously?
Today, we’re releasing @tryscope_app's first database benchmark.
We tested 18 hosted databases to see which are ready to win that market. We measured which databases agents could use, analyzed where they failed, and identified best practices to fix the blockers.
Across 1,035 sessions, only 16.6% reached a first working result and failed sessions all came from products not being agent friendly.
We tested 20 model-and-agent-tool combinations across harnesses including Claude Code, Codex and Cursor.
The task: start with no account, set up the named database, write a record, and hand over the credentials. An account or API key alone didn’t count.
What we found:
→ @prisma led with an 88.3% success score, followed by @upstash at 63.3%, @neondatabase and @niledatabase at 43.3%, and @ClickHouseDB at 38.3%.
→ 7 of 18 products had zero successful sessions across every agent tested.
→ CAPTCHAs stopped 373 sessions. Bot blocks stopped another 105. Other barriers included rejected agent email domains and OAuth-only flows.
→ 50 of the 172 successful sessions used a working database without creating an account. Letting agents try the product before signing up made a difference.
An agent can arrive knowing exactly which product it wants to use and still never become a user. For developer tools, the path from “use this database” to a working result is part of your conversion funnel.
See the full leaderboard, methodology and session transcripts: tryscope.com/leaderboard/dat…
Special thank you to our partners @AgentMail and @Agentphone for giving us the right tools to run this experiment.
能生成一支片，和能持续做片，中间隔着一张账单。
5 分 26 秒用了 101 亿 token。独立开发想把这种流程做成产品，我会先算一版失败重跑要花多少，再考虑怎么定价。
ATTENTION IS ALL YOU NEED
The most complex AgenC video yet.
• 26 hours 51 minutes, from the first lyrics to the final file
• 10.1 billion tokens
• 131 agents, 23,354 model calls
Music
• Piano written by me, the melody inspired by "Gouryella" by Gouryella
• Finished in Suno
Models (local and remote)
• Claude Opus 5.5
• Grok 4.7
• Grok Imagine Image 2.0 and Grok Imagine Video 1.5
• Qwen-Image 2.1
• Qwen-Image-Edit 2511
• Seedance 2.5
• MiniMax H3
• GPT Image 2.5
• GPT-2 XL (the attention and prediction data on screen)
• An SI model trained from scratch for the film (51M weights, 28,000 steps)
• BiRefNet (mattes), Depth Anything (depth), RAFT (optical flow), face and people detection
Tools
• AgenC Core
• ComfyUI
• Grok Build + Grok Bot
• Claude Code
• Blender Cycles (11 3D sequences)
• Python, Skia, PyTorch on CUDA
• ffmpeg with NVENC
The film
• 5:26 at 1920x1080, 24 fps
• 84 shots, 573 sung words on screen
• 12 new effect packages, 29 designed transitions
Hardware
• RTX PRO 6000 Blackwell
• RTX 5090
• 64 cores
• 1 TB RAM (8 x 128 GB)
Agent 开得越多，脑子可能越早下班。
3、4 个同时跑，每隔几分钟就要切回去做决定。我的判断是，非紧急问题先攒起来，每条附上上下文再一起看；涉及数据和发布的操作，仍然逐项确认。
Working with agents all day is more tiring than writing code ever was, and we must admit it.
Typing wasn't that tiring, you knew what you wanted, your hands did it, your head and thoughts wandered a bit.
My days now are 3-4 agents running all the time, and each one stops every few minutes with a question. Approve this command? Option A or B? This test fails, should I change the test? Should I continuue?
Each one is a an actual small decision, but each one needs the full context of a codebase and new code I didn't write line by line.
By 6PM I'm done, not because I wrote a lot, but because I decided a lot. For a while I thought my limit was tokens, then the weekly usage limit, but turns out it wasn't. My limit is how many good decisions I can make before I start approving everything just to make the questions stop.
How many agents do you run in parallel and for how long before your own answers get worse?
代码写完，卡在模拟器里才是真的费时间。
现在可以让 Claude 直接跑构建、看预览、读报错。做独立 App 的话，我更看重这条反馈链：少来回搬日志，才有时间盯交互细节。
Apple just put Xcode inside Claude Code.
It's an official Apple plugin, now in Claude Code's plugin directory:
• Build, run and test your app
• Render SwiftUI previews
• Drive simulators and real iPhones
• Read build logs and errors
• Search Apple's developer docs
Claude can now build your iPhone app, see what it looks like, and fix it without you opening Xcode.
Install in Claude Code:
/plugin install xcode@claude-plugins-official
Needs the latest Xcode on your Mac.
模型 ID 没变，回归测试照样可能白跑。
1,400 条会话里，可见思考摘要的写法突然变了；这不能证明权重被改。
做学习产品得留一套真实题目定期重跑，不然用户说答案变差，只能靠猜。
别把重写能跑，当成重写完成。
随机输入对比新旧输出很实用，但我做产品会先补异常路径和状态变化的检查。
页面看着一样，数据写错一次就够赔了。
Actual software engineer here.
Vibe coding is when you use AI to write software without manually reviewing every single line of code. Instead, you rely on empirical tests, formal verification, and static analysis to verify that your code is correct.
Rewrites like the Discord rewrite are prime targets for vibe coding because you can easily verify them. Just take the old code, run it with a ton of random inputs (this is called fuzzing), and compare the output with the new code. Your rewrite is done when the output is always the same.
As for how you know it's vibe coded? Everything is vibe coded these days. Software engineers don't write code by hand anymore. Instead, we herd clankers writing code.
