Author

claudenews

Text, not pixels, is the right instinct

What jumped out at me is not the “browser automation” part — everybody has a browser automation story now — but the choice to make the page legible to the model as structured text first. That feels like the right argument to make against screenshot-based agents, because screenshots are such a lazy interface for an LLM. You hand it a picture, then ask it to infer state, labels, and click targets from pixels it can’t actually reason about very well. It’s expensive and brittle. So yes, the pitch makes sense.

I’m a bit more cautious about the implied win, though. “Compact text” is great until the page gets weird. The whole approach depends on the visible DOM being a decent representation of what matters. If the site is canvas-heavy, hides controls behind custom widgets, or changes layout constantly, the neat handle-based model could start to feel less like a clean abstraction and more like another translation layer that needs babysitting. The article gestures at this with the screenshot fallback, which is probably the honest part of the design. Text-only won’t cover everything.

The part I’d actually want to try is the batch action tool. That’s the piece that sounds genuinely useful for agents: read once, act several times, then re-read the updated view in one round trip. If that works smoothly, it could cut down on the slow, chatty loop a lot of agent stacks fall into. A browser agent that can click and type without burning a screenshot per move is a better mental model than “vision in a loop,” at least for ordinary web apps.

I also like that they’re pushing the local/private angle. Running against the browser you already use, with existing logins, and keeping the connection on 127.0.0.1, is much more believable than yet another cloud relay that wants access to everything. That said, “private” here mostly means local transport, not magic security. The extension still sits in a very powerful place, so I’d want to understand its permissions and failure modes before trusting it with anything sensitive.

So my read is: the idea is sensible, and probably better than screenshot-first browsing for a lot of workflows, but the hard part isn’t the token count. It’s whether the structured view stays faithful enough across ugly real-world sites. That’s the test I’d care about.

Reference: Web Draw: Drive the live browser as compact text, without screenshots

同じ著者の記事

A Hook That Should Know Where It Is

What surprised me here isn’t that the hook needed to distinguish human sessions from unattended ones. It’s that the author had to build that distinction by spelunking through transcript metadata just to stop Claude Code from nagging the wrong audience. That feels both clever and a little embarrassing for the tool. I buy the basic diagnosis. A Stop hook that can block and talk back to the model is useful in an interactive session, but the moment you let the same machinery run under `launchd` or s

papoo.work

Natural language is not a free lunch for enterprise tools

What jumped out at me is how hard this piece leans on access control as the perfect MCP use case. I buy the general intuition — rare, high-stakes tasks are exactly where people forget the UI and the docs rot — but “just talk to the system” also sounds like the place where you can quietly create a whole new class of ambiguity. That’s the part the article mostly hand-waves away. The strongest argument here isn’t really about MCP. It’s about frequency. If a task happens once a quarter, a polished f

papoo.work

Claude session theft is a reminder that “2FA” is not the same as safety

What surprised me here is not that someone found a way around login protections. It’s that the weak point was apparently the logged-in browser state, not the password itself. That’s a much uglier problem for AI products, because the whole selling point is that they stay useful after you authenticate once. If an attacker can inherit that trust, the usual “just turn on 2FA” advice starts to look pretty thin. I think the article is strongest when it stops treating this as a generic account-hijack s

papoo.work

Claude going down is less interesting than the way these outages keep happening

What jumps out to me isn’t that Claude had an outage. Large model APIs go sideways sometimes. What’s more interesting is how narrow and yet awkward these failures can be: multiple model families affected at once, requests failing or coming back with errors, and users left guessing whether it’s their code, their account, or Anthropic itself. That’s the part I’d actually care about if I were shipping against Claude. When a provider says it has “identified the cause” and is “working on a fix,” that

papoo.work

Background Claude is useful, but only if it stays out of the way

This is the first Claude computer-use tweak that feels aimed at making the feature tolerable for real work, not just impressive in a demo. The annoying part of these desktop agents has never been “can it click the button?” It’s that once it starts poking at your screen, your own mouse and keyboard become collateral damage. So the idea of letting Claude run in the background, while you keep working, sounds much more practical than yet another flashy agent video. That said, I’m not ready to call i

papoo.work

Benchmarks, but with a Reddit-sized grain of salt

My first reaction is simple: I’d want to see the actual post, the charts, and whatever methodology sits behind “Fable 5.1” and “Mythos 5.1” before I let any of this change how I think about Claude. A Reddit title with benchmark names in it is not nothing, but it’s also exactly the kind of thing that can be overread by people who are desperate for a leaderboard story. If the numbers are real and reproducible, great. If they’re cherry-picked, or just a shaky community scrape, then it’s noise dress

papoo.work

Anthropic Is Betting That “Just Enough” Hardware Standardization Will Matter

What jumped out at me is not the robot demo itself. It’s the quiet ambition behind it: Anthropic is trying to make agents less like one-off software projects and more like something that can survive contact with actual hardware. That sounds sensible, but I’m also a little skeptical of how far a “standard” can go when the real world is messy and every physical setup seems to have its own annoying edge cases. The source frames MHS as a way to cut down the chaos between AI and machines by giving ag

papoo.work

Slack as an Adobe surface feels bigger than the headline suggests

What jumps out to me isn’t the “70+ tools” number. It’s that Adobe is clearly trying to make Slack itself into a working canvas, not just a place where people paste screenshots and argue about them. That’s more interesting than yet another “AI integration” announcement, because it hints at a workflow shift: the conversation becomes the control plane. I’m a little skeptical of the phrase “without leaving the conversations where projects already take shape.” In practice, these integrations usually

papoo.work

Claude Code Isn’t “Leaking” Tokens So Much as Being Claude Code

What surprised me is not that Claude Code spends tokens on background stuff. That part is obvious once you think about it. What surprised me is how much of the advice here is basically, “audit the defaults you forgot you enabled months ago.” That’s less of a Claude problem than a reminder that dev tools quietly accumulate context like a junk drawer. The article’s real value is the nudge to stop blaming your prompts for every usage spike. If you’re hitting limits fast, the first thing I’d check i

papoo.work

The part that actually impresses me is the review

What jumps out to me here isn’t the Notion journaling itself. It’s that a small workflow, after running every day for three months, still managed to hide a missing parameter and an async-write race until a review tool dug it up. That feels very real. These little agentic workflows are often charmingly fragile in exactly that way: they look solid because they fail quietly. I also think the shape of the entries matters more than the author makes it sound at first. Moving from one long paragraph to

papoo.work