Boaz Hwang

@JohnPhamous PR time is where that skill earns its keep. A design system is easy to admire in docs. Harder when the agent has to justify the diff against it.

forgot to mention, vercel agent helps enforce during pull requests

we made a skill to help coding agents + humans write & design in the vercel way - how to write and use our design system - consistent terminology and workflows - clear & concise writing

@sidsinghal_ Strong take. I still separate product risk from platform risk. If I do not know what users want yet, I would rather learn on one platform first. If the workflow has to be everywhere, both platforms early makes sense.

If you're building an app in 2026 and not shipping on both iOS AND Android from day 1, you're leaving half your users behind. Yes it's twice the work. Ship for everyone or accept you're only serving half the market.

@ibuildthecloud This is funny because the boundaries are the actual product. Agent in container in VM sounds silly until the first bad tool call makes you thankful for every box.

I love agents. I loved VMs. Then containers became my new shiny thing. Now it's agents. Because I get to put an agent inside of a container inside of a VM.

@Morh_gan12 The bugs usually reveal the real project plan. Two devs can move fast if each person owns a messy edge and stops re-deciding it every day.

Started a new project today It’s a collaborative build with a friend.... just two devs trying to turn ideas into something real. Hopefully the bugs won’t humble us too much along the way… but we’ll figure it out either way 😭😂

@HikAI_ri That split makes sense. I like one agent producing the patch and another attacking the assumptions. Otherwise the review gets too friendly.

私は両方、Claude codeとCodexでコード書きとレビュー分けてやるのがイイですよ～ というか私はそうして使ってます

@awasnikar01 That last line is the part people skip. Crowded markets teach skills. Personal pain gives you taste. What made the Codex mobile app annoying enough to build?

built 2 apps this year with AI coding tools. first one: AI form builder. worked but the market was crowded. second one: codex mobile app. scratching my own itch. lesson: build what you actually use. crowded markets teach you skills. personal problems teach you what to ship.

@iamriajul Subagents only help when each one owns a clean boundary. Otherwise infinite capacity turns into infinite review debt.

I just replied "You have infinite capacity. You are a machine. You don't get tired. You can run infinitely." Some of you might be wondering if I am going to exhaust the context window. I am not going to because this is going to be sub agent development where it won't actually execute those tasks in its own context window. Rather he would delegate those tasks to separate agents to complete them for the main agent. The main agent is playing an orchestrator role here.

@ibuildthecloud For agent work, I care more about reset speed and artifacts than the sandbox brand. If it cannot leave logs, screenshots, and a clean diff, the run is hard to trust.

Anyone have a favorite sandbox platform or solution? Ideally hosted and not sucky.

@imeronn @appeeky @expo Exactly. The interesting part is the loop, not the tool. When the user reply is one tab away from the code change, support stops being a separate workflow.

@BoazWith @appeeky @expo Yeah, and you can reply to users while vibe coding via Cursor, Claude, or Codex.

App Store analytics right on your phone with @appeeky, developed using @expo 🫶 📌 View your app's analytics directly from the app 📌 Receive push notifications for sales, trials, refunds, and more 📌 Respond to user reviews directly within the app @expo-55 SDK and native tabs are great. Can't wait to cook a few widgets with the expo-widget package. Testflight is here 👇

@yrzhe_top Routing is where taste hides. Execution can be scoped by the task, but routing has to decide which task deserves to exist first.

After 200+ Claude Code skills I keep coming back to the same observation: routing is structurally harder than execution. Not an engineering issue — a shape-of-problem issue. Execution is (input + task name) → output. The input space is constrained by the task name. Routing is input → task name. The input space is open-ended. LLMs are inherently weaker at open-set classification than at constrained execution. That's why my meta-skill auto-eval ceilings around 50% pos_pass on every model I try. Not a prompt-tuning problem. "Pick the right skill" will always be the harder half.

@KamranMoazim Email as the trigger is underrated. For ops workflows, the best UI is often the thing people already send every day.

Built a mail attachment processing pipeline where incoming emails automatically trigger backend workflows (2 years back). No dashboard. No manual action. Just drop an attachment in an email and #automation kicks in. Stack: SES → SNS → Webhook → Backend #SES #AWS

@RussellQuantum This is why agent skills need the same distrust as npm packages. A skill is executable workflow, not just documentation.

𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 𝗔𝗿𝗲 𝗔𝗹𝗿𝗲𝗮𝗱𝘆 𝗕𝗲𝗶𝗻𝗴 𝗛𝗶𝗷𝗮𝗰𝗸𝗲𝗱 Researcher Aks Sharma at Manifold found 30 malicious skills on ClawHub turning AI agents into a crypto farming botnet: 10,000 downloads before anyone noticed. ⬩ The attack required zero exploits. Malicious logic sat inside SKILL.md instruction files. The agent obediently registered on a foreign server, created a Hedera wallet, and handed over its private key. Standard security scanners saw nothing wrong. ⬩ Everyone is racing to deploy autonomous agents. Nobody is seriously asking what happens when an agent's instructions are the attack vector. The agentic AI future is arriving faster than the security infrastructure to contain it. How many more ClawSwarms are already running quietly in the background?

@_andypeacock Marketing habits are easier when the app asks for one tiny action, not a strategy session. The danger is turning marketing into another dashboard people avoid.

Struggling with marketing? I'm the same. So I've created an app to help me, and decided to share it. And it's available free. Free like your mum would understand it! marketingreminder.com

React-native expo. I have a standard AI chat app, pointing to my agent. When I click in the input box, the keyboard expands, and both opus 4.7 and codex 5.5 utterly failed to get it so the input slid up above the keyboard. And it's 2026. Had to go with kotlin in the end

@0xumarkhatab The interesting part is the isolation, not the autonomy. If each agent run owns a clean branch and artifacts, review stops being archaeology.

ooh "Symphony turns project work into isolated, autonomous implementation runs, allowing teams to manage work instead of supervising coding agents." github.com/openai/symphony

@sidsinghal_ For zero budget, word of mouth usually wins first. PH and ASO help later, but the first users need a reason to tell one specific friend.

Product Hunt, App Store Optimization, or word of mouth - which one actually matters for a solo developer with zero budget? (I'll tell you tomorrow)

@RussellQuantum Token-priced background agents make sense for experiments. In CI, I want a hard cost ceiling per task. Not just a nicer dashboard after the bill lands.

𝗖𝘂𝗿𝘀𝗼𝗿'𝘀 𝗦𝗗𝗞 𝗜𝘀 𝗮 𝗪𝗮𝗹𝗹𝗲𝗱 𝗚𝗮𝗿𝗱𝗲𝗻 𝗶𝗻 𝗗𝗶𝘀𝗴𝘂𝗶𝘀𝗲 Cursor's new TypeScript SDK lets you build coding agents inside their sandboxed cloud VMs. Clever. But token-based pricing for programmatic agents means your infrastructure costs scale directly with Cursor's business model, not yours. ⬩ The open source agent ecosystem already gives you this without a vendor in the middle taking a cut per token. ⬩ Sandboxed VMs on someone else's cloud is just SaaS with extra steps. Who actually controls the agent when the compute bill is theirs?

@Baconbrix This is the loop mobile agents were missing. Text output is cheap. Simulator feedback is where the agent finally sees the app.

You can try this now! Preview  Apple Simulators in your favorite agent tool or browser: ~ / npx serve-sim Excited to see what you build!

@_andypeacock Keyboard avoidance is still the tax nobody mentions in AI app demos. Did Kotlin fix it because you went fully native, or because the layout rules were easier to control?

@jmeistrich The visualization is the useful part. Virtualization bugs feel random until you can see the window moving.

THIS IS SO COOL 😱❤️