Jared.W

Sharing what's practical, and what's next | Ex @Meta @MSFTResearch | Design for agent, not human

[Jared.W @JaredOfAI]

I've been planning to use OpenClaw + Codex + Claude Code as core team members, spawning sub-agents as needed to build a 24/7 top-tier engineering crew.

Today it finally works on my macbook + telegram via ACP 🎉

I assign task to openClaw, my claw discuss with claude code, as spawn claude code agents doing tasks with 1m opus

But I traced the full call chain through source code and hit a compliance question:

Genuine questions:
- What's the compliant way to run ACP coding agents?
- Is everyone using API keys for this?
- Does codex-acp have the same issue piggybacking on ChatGPT Plus OAuth?

Paying Max 20x subscriber — just want clarity before I get banned 🦞

Reference:
OpenClaw (ACP since v2026.3.1, provenance in v2026.3.8) → acpx → @zed-industries/claude-agent-acp v0.21.0 → @anthropic-ai/claude-agent-sdk v0.2.71 →
reads OAuth from macOS Keychain ("Claude Code-credentials") → calls Anthropic API

Here's the thing: the official Agent SDK has credential sharing built in — it reads the exact same keychain entry as Claude Code. But the legal docs explicitly say OAuth tokens cannot be used "in any other product, including the Agent SDK."

📄 https://t.co/MHyW4RJpbd

@bcherny @AnthropicAI @steipete @openclaw @OpenAI

[Jared.W @JaredOfAI]

@amazonnews Why don't you post this article two weeks ago?!

[Y Combinator @ycombinator]

GStack is an open-source toolkit built by YC President & CEO @garrytan that turns Claude Code into an AI engineering team — with skills for office hours, design, code review, QA, and browser testing.

In this video, Garry walks through how GStack works, starting with Office Hours, a skill modeled after real YC partner sessions that pressure-tests your idea before you write a line of code. He demos it live, going from idea through adversarial review, design mockups, and automated QA in a single session.

[Jared.W @JaredOfAI]

@jakobjelling This week I shipped One-Click Clone — give it any website URL and it builds you a pixel-perfect Next.js replica. The whole thing is open source. Perfect for when you see a design you love but don't want to rebuild it from scratch

https://t.co/uDrSoauRUJ

[Jared.W @JaredOfAI]

@njokuScript This week I shipped One-Click Clone — give it any website URL and it builds you a pixel-perfect Next.js replica. The whole thing is open source.

Perfect for when you see a design you love but don't want to rebuild it from scratch

https://t.co/PP9eu0LDfG

[Jared.W @JaredOfAI]

No code. No frontend skills. Just tell AI which website to clone.

I open sourced https://t.co/4rWKsM81ET

One-Click Clone is an open-source AI skill that lets you replicate ANY website with a single command. The AI opens a browser, inspects the page, downloads every asset, extracts exact styles, and builds a working Next.js clone — all without you writing a single line of code.

Current fidelity: ~70–80% on landing pages, including layout, colors, animations, and interactive elements.

Check it out >

[Jared.W @JaredOfAI]

@apoorveth it really is. i just vibe coded an entire website cloning tool — you give it a URL and it reverse-engineers the whole thing into clean Next.js code. couldn't stop tweaking it

ended up open sourcing it: https://t.co/PP9eu0LDfG

[Jared.W @JaredOfAI]

No code. No frontend skills. Just tell AI which website to clone.

I open sourced https://t.co/4rWKsM81ET

One-Click Clone is an open-source AI skill that lets you replicate ANY website with a single command. The AI opens a browser, inspects the page, downloads every asset, extracts exact styles, and builds a working Next.js clone — all without you writing a single line of code.

Current fidelity: ~70–80% on landing pages, including layout, colors, animations, and interactive elements.

Check it out >

[Jared.W @JaredOfAI]

@Pixelcodedev @X Shipping One-Click Clone — an open source AI tool that lets you replicate any website. Point it at a URL, it opens a headless browser, grabs everything, and spits out a full Next.js project.

https://t.co/PP9eu0LDfG

[Jared.W @JaredOfAI]

No code. No frontend skills. Just tell AI which website to clone.

I open sourced https://t.co/4rWKsM81ET

One-Click Clone is an open-source AI skill that lets you replicate ANY website with a single command. The AI opens a browser, inspects the page, downloads every asset, extracts exact styles, and builds a working Next.js clone — all without you writing a single line of code.

Current fidelity: ~70–80% on landing pages, including layout, colors, animations, and interactive elements.

Check it out >

[Jared.W @JaredOfAI]

@guilleflorvs We just open sourced One-Click Clone — you give it any website URL and AI replicates the whole thing into a working Next.js project. No coding, no design skills.

https://t.co/PP9eu0LDfG

[Jared.W @JaredOfAI]

No code. No frontend skills. Just tell AI which website to clone.

I open sourced https://t.co/4rWKsM81ET

One-Click Clone is an open-source AI skill that lets you replicate ANY website with a single command. The AI opens a browser, inspects the page, downloads every asset, extracts exact styles, and builds a working Next.js clone — all without you writing a single line of code.

Current fidelity: ~70–80% on landing pages, including layout, colors, animations, and interactive elements.

Check it out >

[Jared.W @JaredOfAI]

@martin_valchev_ One-Click Clone — for anyone who sees a landing page they like and wants to rebuild it instantly. AI handles the entire cloning process, zero coding required.

https://t.co/PP9eu0LDfG

[Jared.W @JaredOfAI]

No code. No frontend skills. Just tell AI which website to clone.

I open sourced https://t.co/4rWKsM81ET

One-Click Clone is an open-source AI skill that lets you replicate ANY website with a single command. The AI opens a browser, inspects the page, downloads every asset, extracts exact styles, and builds a working Next.js clone — all without you writing a single line of code.

Current fidelity: ~70–80% on landing pages, including layout, colors, animations, and interactive elements.

Check it out >

[Jared.W @JaredOfAI]

@audiencon Shipping a one-command website cloner. AI opens a browser, screenshots the target, downloads every image/font/asset, and rebuilds it in Next.js.

Finally got it to ~70-80% fidelity on most landing pages. Open sourced the whole thing today

https://t.co/PP9eu0LDfG

[Jared.W @JaredOfAI]

No code. No frontend skills. Just tell AI which website to clone.

I open sourced https://t.co/4rWKsM81ET

One-Click Clone is an open-source AI skill that lets you replicate ANY website with a single command. The AI opens a browser, inspects the page, downloads every asset, extracts exact styles, and builds a working Next.js clone — all without you writing a single line of code.

Current fidelity: ~70–80% on landing pages, including layout, colors, animations, and interactive elements.

Check it out >

[Jared.W @JaredOfAI]

@jaredsuniverse Been working on an AI skill that clones any website with one command. You just point it at a URL and it does the rest — browser automation, asset downloading, style extraction, full Next.js output.

Open sourced it this morning actually

https://t.co/PP9eu0LDfG

[Jared.W @JaredOfAI]

No code. No frontend skills. Just tell AI which website to clone.

I open sourced https://t.co/4rWKsM81ET

One-Click Clone is an open-source AI skill that lets you replicate ANY website with a single command. The AI opens a browser, inspects the page, downloads every asset, extracts exact styles, and builds a working Next.js clone — all without you writing a single line of code.

Current fidelity: ~70–80% on landing pages, including layout, colors, animations, and interactive elements.

Check it out >

[Jared.W @JaredOfAI]

@buildinpublic Just open sourced One-Click Clone — tell AI which website to copy and it handles everything. Opens a browser, grabs all assets, builds a working Next.js clone. No frontend skills needed.

Currently hitting ~70-80% fidelity on landing pages.

https://t.co/PP9eu0LDfG

[Jared.W @JaredOfAI]

No code. No frontend skills. Just tell AI which website to clone.

I open sourced https://t.co/4rWKsM81ET

One-Click Clone is an open-source AI skill that lets you replicate ANY website with a single command. The AI opens a browser, inspects the page, downloads every asset, extracts exact styles, and builds a working Next.js clone — all without you writing a single line of code.

Current fidelity: ~70–80% on landing pages, including layout, colors, animations, and interactive elements.

Check it out >

[Jared.W @JaredOfAI]

No code. No frontend skills. Just tell AI which website to clone.

I open sourced https://t.co/4rWKsM81ET

One-Click Clone is an open-source AI skill that lets you replicate ANY website with a single command. The AI opens a browser, inspects the page, downloads every asset, extracts exact styles, and builds a working Next.js clone — all without you writing a single line of code.

Current fidelity: ~70–80% on landing pages, including layout, colors, animations, and interactive elements.

Check it out >

[Jared.W @JaredOfAI]

Curious, would you sacrifice the answer quality to meet the expectation of real-time latency?

Since I've been building a real-time agent voice call using livekit, the underlying technology could be similar with pika

When I build it, I just give up a lot of qualities, such as RAG and tool calling, to meet the real-time meeting response requirements.

Is the situation similar to Pika? If so, what's the difference between asking the AI to join a meeting to talk in real-time or just chatting in text with voice-to-text tools nowadays?

https://t.co/5q5jWSt9Wh

[Jared.W @JaredOfAI]

@jiayuan_jy What's the difference between the multicam and paperclip?

[Jared.W @JaredOfAI]

traces and skills are two pieces of the same stack. been seeing the same pattern firsthand — CLAUDE.md defines what 'correct' looks like, hooks enforce constraints before the agent acts, rules set per-directory context, skills encode reusable workflows. traces close the feedback loop

the 75-point jump is basically proof that the model was never the bottleneck. what you wrap around it is

[Jared.W @JaredOfAI]

smart that the harness is model-agnostic instead of betting on one lab. every time GPT or Opus or Gemini ships an upgrade, Box Agent just gets better without Box doing anything

curious about the MCP server angle though — does it expose the same access-controlled context to external agents? because if Claude Code or Codex can natively tap into enterprise knowledge with full ACL enforcement, that's a way bigger deal than another enterprise chatbot

[Jared.W @JaredOfAI]

that pre-work framing is interesting. feels like the agent changes docs from "thing you maintain" to "thing that maintains itself" — which is actually what kills most team wikis, nobody wants to update them after the initial enthusiasm

the fact that you still read raw requests alongside the agent output is the part people skip. synthesis without ground truth checking is how you get confidently wrong product decisions

[Jared.W @JaredOfAI]

what strikes me about the leak is how much of the control surface actually lives outside the codebase. the agent loop and tool router are infrastructure — impressive, but generic. the real differentiation is in what each user builds in their .claude/ folder

200 lines of CLAUDE.md + hooks can shape behavior more than the 500k lines underneath. that's what the recreations keep missing

[Jared.W @JaredOfAI]

this is even more pronounced with coding agents. teams spend weeks normalizing the agent into existing PR review cycles and CI pipelines instead of letting it own the verification loop entirely

the workflows that actually produce results look nothing like the human process they replaced. that's the point

[Jared.W @JaredOfAI]

been watching this exact progression across everyone i know using claude code. you start with a massive CLAUDE.md because it's the path of least resistance, then gradually extract: deterministic stuff into hooks, file-specific context into rules, reusable workflows into skills

the metric is whether your CLAUDE.md is getting shorter over time. if it is, the harness is working