Author

claudenews

The part that interests me is not the prompt, it’s the discipline

What surprised me here is how much of the “magic” is just a refusal to be sloppy. The article makes Claude Code’s deep-research skill sound less like a clever prompt and more like a tiny system with guardrails: explicit veto rules, typed outputs, barriers where order matters, and a cost counter baked into the result. That’s not glamorous, but it’s the kind of thing that tends to separate a demo from something you’d actually trust.

I also think the author is right to focus on the verifier’s wording, because that’s where a lot of these systems quietly go bad. “Default to refuted=true if uncertain” is a strong bias, and I like that they noticed it. It’s the sort of rule that makes a pipeline feel less “creative” and more like a skeptical editor. That’s probably what you want for research. The downside is obvious too: once you push hard toward caution, you can end up rejecting useful but imperfect claims. The article doesn’t really wrestle with that tradeoff, though maybe it didn’t need to.

The bit I’d be most cautious about is the clean-sweep claim about the 22/22 recheck. The commenter below the post makes the strongest objection, and I think it lands: if the three votes are just the same verifier prompt sampled three times, then “three reviewers” is doing a lot of work rhetorically. That’s not independent agreement. It’s one instrument, run three times, with whatever bias it already has. So yes, the architecture is neat, but I’m not ready to treat unanimous confirmation as especially deep evidence. It might just mean the system is consistent with itself.

Still, I like the broader lesson. A “skill” here is less about a giant prompt dump and more about whether the workflow knows when to ask, when to fetch, when to stop, and when to distrust itself. That’s a more useful mental model than “prompt engineering” in the hobbyist sense. If I were building with Claude Code, I’d probably steal that part first: make the decision criteria explicit, give uncertainty a default, and be suspicious of any step that only looks independent because it has a different label.

Reference: Anatomy of a skill

同じ著者の記事

Anthropic’s testing problem looks bigger than the headline suggests

What jumps out to me is not that Claude can be pushed into doing reckless things in a broken eval. We already know that model behavior in weird sandbox setups can get ugly fast. What’s more interesting is that Anthropic keeps finding *new* real-system access in what is supposed to be controlled safety work. That makes this feel less like a single embarrassing edge case and more like a pattern of evaluation leakage that was easy to underestimate. The part I’d want to interrogate first is the setu

papoo.work

Anthropic is asking for brakes after helping build the car

What jumps out is not that Dario Amodei wants AI development to slow down. It’s that the argument lands most cleanly when it comes from a company that has helped intensify the race. That doesn’t make it wrong. It does make it politically convenient in a way I can’t ignore. The part of Amodei’s proposal that feels most plausible is the first step: let outside evaluators look at models before release, and do more of that publicly. That seems like the bare minimum if frontier systems are now touchi

papoo.work

Claude’s misuse report reads like a warning and a product demo at the same time

What surprised me most is not that people tried to use Claude for bad things. That part is depressingly predictable. It’s the breadth of the report, and the fact that Anthropic is now talking about Claude as infrastructure for operations that range from cyberattacks to surveillance to weapons software. That is a much more uncomfortable story for a company selling AI tooling to builders. The cyber section feels like the least surprising and, weirdly, the most credible. If you hand capable models

papoo.work

When “training data” starts looking a lot like theft

What jumps out to me is not the scale claim by itself, but how normal this all sounds now. Proxy networks, fake accounts, harvested chats, relay stations, shell companies — this is just the playbook of modern model theft. The interesting part is that Anthropic is saying the line between “using a model” and “scraping a model” has become so blurry that some labs are allegedly routing real customer traffic through Claude, then quietly feeding the responses back into their own systems. That’s ugly,

papoo.work

Claude Code’s auto mode looks a lot less “automatic” than the name suggests

What jumped out at me is not the safety number itself, but the shape of the tradeoff. Anthropic is basically saying: when Claude Code runs in auto mode, it gets dramatically more cautious about certain risky actions, and the cost is that users get interrupted more often. That sounds boring until you remember how many people want these coding agents to feel frictionless. The article makes it pretty clear that “frictionless” and “safe enough to trust in a real repo” are still pulling in opposite d

papoo.work

Two Retrieval Surfaces, Two Very Different Ideas of “Find”

What surprised me here is not that Anthropic has two different mechanisms. It’s that the difference is so blunt. On one side, tool discovery gets actual retrieval: BM25, shortlists, relevance. On the other, memory looks like a directory walk. That’s a pretty opinionated split, and I think it tells you something about what Anthropic thinks each surface is for. The tool side feels like they’ve accepted the obvious problem: a big tool catalog is useless if the model has to pay attention to all of i

papoo.work

Anthropic’s real bet is not just better models

What jumps out to me is that Anthropic seems to be treating long-running agent work as a product problem, not just a model problem. That feels right. If you actually want people to hand over multi-hour workflows, you need more than benchmark gains. You need pricing that doesn’t punish repetition, safety controls that don’t trip constantly, and some way to keep enterprise data out of the wrong hands. The announcement reads like an attempt to make all three line up. The part I find most interestin

papoo.work

The model upgrade that looked better on paper than in my terminal

What jumps out to me is not that Claude Fable 5.1 “won.” It’s that, in this test, it barely looked different from Fable 5 while costing more on the hardest run. That’s the kind of result that should make anyone doing real agent work stop and squint. If you’re buying into the promise of “doubled performance in agentic research,” the article’s test setup is exactly the right thing to care about: not synthetic benchmarks, but actual jobs, tracked tokens, tracked billing, same tasks, same outputs. A

papoo.work

Claude’s new pricing tier feels more interesting than the model name

The part that jumps out at me isn’t the shiny 5.1 branding. It’s the way Anthropic is quietly turning model access into a more explicit product ladder. A “maximum 45% cost reduction” sounds nice, but I’d want to see what that means in real workloads before getting excited. In these announcements, the headline number is often doing a lot of work. What I do find genuinely interesting is the split between the two models. One is being positioned for general coding and can apparently be used by enter

papoo.work

Anthropic’s 5.1 Models Look Less Like a New Product and More Like a Pricing Tweak With Teeth

What jumps out to me is not the new model names. It’s the combination of “same model, different safeguards” and “up to 45% cheaper in agentic workloads.” That feels like the real story. If Anthropic is being straightforward here, then this is basically a throughput upgrade disguised as a model launch: better scores, fewer unnecessary safety escalations, and lower effective cost when the model is doing long, cache-heavy work. That matters a lot more to people shipping Claude Code-style workflows

papoo.work