The anthropics Skill | Claude’s Official Skill Library

[0:00]Hedge. This is Liam in for bear. Today we tear down a skill that isn't about making a brutalist reel. It's about reading a whole company's output, code, papers, content, capabilities, and reporting from your own bench, not from the press release. It's called anthropics. Here's how it actually works. The easy read of a skill named for company is that it's another clawed tutorial channel. It isn't. It's a beat journalism sense that reads the artifact, not the docs. Everything the skill does exists to keep it honest when the docs would rather do the explaining for it. A skill is a folder Claude reads first. The enthropic skill is small on

[0:33]Disk. One skill file 142 lines sitting under skills/make. What's big is what the file references. Two sibling engines, four modes, and one hard disqualifier that every episode has to pass. The doctrine is short because the discipline is what does the work. One router, four modes. Dashreo dispatches to get explainers analysis engine. Dash paper dispatches to AI paper. Dash content asks what has changed since the material was published. Dash capability runs the thing and shows the real output often side by side against a competitor. That last mode is the whole reason this skill exists. No vendor channel will ever ship it. First design decision and it's the whole thesis. Quote from the

[1:12]Skill file. If the episode could be made by reading the docs, it is not the series. Every episode needs at least one spine beat only visible by opening the artifact and running it. Point at that beat or the episode is not ready. The skill will report the gap rather than build around it. That single rule kills the 90% of AI content that is downstream of the same three landing pages. Second decision, structure and history come free from a repo. Behavior is the thesis. So the skill file lists reusable probes. Every URL contacted every file read outside the working directory. The default value of every safety and

[1:43]Privacy feature because quote a feature that exists but ships off tells a different story than the paragraph describing it. For Dash paper, the sharpest fair question is, does the release code reproduce the published claim? The honest answer is often partially, and that is a more interesting finding than either open science tick or sham. Third decision register is a machine of its own. Another perspective, not a correction, never they're wrong or they're behind. Independent stated once plainly, no hedging. Builds with Claude Daily, nobody's paying for this. Situated in a date and a real task, so the episode becomes history instead of going stale. and a matching claim is a finding some

[2:20]Repos are exactly what they say. And manufacturing a gap where none exists is the dishonesty this beat is positioned against. Here is where the skill bites. It leans on two sibling engines, git dash explainer and AI- paper that aren't shipped inside this toolkit snapshot. Point dash repo or dash paper at a target and the router has nowhere to dispatch. The fix is to install those skills not to fake the analysis. Second bite, the generation honesty law. Never generate an image that poses as evidence of a fact. That's a hard ceiling. It rules out most of what makes a video pop. And the skill accepts the trade

[2:52]Because the credibility is the whole product. The verdict, the anthropic skill is a beat, not a topic. Four modes, one router, one disqualifier, one register. It reads the artifact rather than the artifacts press release. And it says so out loud. Its cost is depending on siblings that have to be present. And its cheap code is a comparison shot no vendor channel can ever ship. Your turn. Paste this into Claude code inside a clone brutalist toolkit where you also have the git dash explainer and AI-Paper skills installed. Use the anthropic skill and dash capability mode to compare Claude and one other model on a real task from your own work. A real

[3:27]File, a real prompt, a dated bench log. Then check the three questions, the skill file names. Is there at least one spine beat visible only by running the artifact? Is every claim cited to a file line or an on-screen output? Did a matching claim survive? If any answer is no, don't build the episode. Report it instead. That was the anthropic skill. Lay them in for bear.