Vollständiges Transkript anzeigen (1.647 Wörter)
Anthropic just accidentally showed everyone the future of work. Their own finance team, the people who report the numbers to the board, quietly built over 70 custom Claude skills, and their CFO went on the record saying one of them now produces the company's entire monthly financial review, 90 to 95 percent finished, before a human even looks at it. Their legal team did the same thing, so did engineering. Here's the strange part.
Almost nobody outside Anthropic uses skills this way, and most people don't even know what they are. So I read every blog post, every doc, and the CFO's full interview and pulled out exactly how Anthropic's own employees use Claude skills and the simple way you can copy each one. Stick around, because the last pattern is the one quietly saving people 10 to 20 hours a week.
First, forget what the name makes you assume. A Claude skill is not a prompt, and it's not a setting. Anthropic's own definition is this: a skill is an organized folder of instructions, scripts, and resources that Claude discovers and loads on its own, exactly when a task needs it. Their favorite analogy, and it's the one that makes it click, is that building a skill is like writing an onboarding guide for a new hire. Think about that.
A prompt is something you re-type every single time for one conversation. A skill is written once, from then on, Claude already knows your format, your rules, your process, the way a trained employee would. You stop re-explaining yourself. That's the whole shift. The people getting huge results from Claude aren't writing better prompts. They've stopped writing prompts for repeated work entirely and started writing onboarding guides their AI keeps forever.
Obvious question: if Anthropic's finance team has 70 of these, doesn't that overload Claude? This is the clever part, and it's worth 30 seconds. Skills load in three levels. Level one: at startup, Claude only sees each skill's name and one-line description, about 100 tokens each. That's it. 70 skills cost almost nothing. Level two: the moment your request matches a skill, Claude opens its instruction file, usually under 5,000 tokens, and reads the full procedure. Level three: only if it actually needs them, it pulls in extra files or runs bundled scripts where the code never even enters the conversation, only the result does. Anthropic calls this progressive disclosure. In plain English, you can install a 100 skills and each one stays invisible and free until the exact second it's useful. That is why a single person can quietly carry an entire department's worth of expertise inside one Claude.
Let's start with the team that shocked me most: Finance, on the Invest. Like the best podcast, Anthropic's CFO, Krishna Rao, said they now have a library of over 70 finance skills that everyone on the team can access through one shared repository. One of them is called the MFR skill, monthly financial review. In his words, it can produce their monthly financial review 90 to 95% ready. Let that land. The most scrutinized document a finance team produces, nearly done, automatically. A weekly revenue and compute report that used to take hours now takes 30 minutes. And in a separate post, Anthropic's own corporate finance lead, Alice Fong, said Claude gave her back 10 to 20 hours every week. Her line is the thesis of this whole video: "Claude holds the integrity layer underneath the work, so my time goes to the narrative on top." Here is how you copy it: Pick the one report you rebuild every week or month, write a skill that spells out the exact sections, data sources, and format. Now you review and add judgment—instead of rebuilding from scratch.
Next, the legal team. And this one is even more clever. Anthropic lawyer, Mark Pike, described writing a separate skill for each practice area: one for employment law, one for commercial agreements, one for privacy. His quote: "A skill can teach Claude to write like an employment lawyer, or format a memo the way a product lawyer would." To get the voice right, they had Claude study 10 of his own past memos until it matched his style. But the best example is this: they built a self-service review tool and pinned it as a site in Slack. Anyone in the company can run their marketing copy or a contract through a skill loaded with the legal team's real guidance, and it flags each issue as low, medium, or high risk with the fix. Do you have rights to this logo? Is this statistic accurate? This trademark needs a symbol. Turnaround dropped from two or three days to about 24 hours. Here is how you copy it: take the rules you repeat to people constantly, your brand guidelines, code standards, your hiring bar, put them in a skill that reviews any input and returns low, medium, or high risk with the fix. You just cloned your own judgment and handed it to your whole team.
Step back, and you see the pattern across every team. Nobody at Anthropic is winning by typing cleverer prompts. They are taking the procedures locked in their heads, the way a report is built, the way a memo is written, the rules they enforce, and encoding each one once as a skill. And you are actually already using this. When you ask Claude on the web to make a PowerPoint, an Excel model, a Word doc, or fill out a PDF, it is silently running Anthropic's pre-built document skills behind the scenes. Skills are not a niche power-user feature; they are the layer that turns a general assistant into a specialist who already knows your job. So let me show you how to build your first one. It is genuinely about five minutes.
Here is the entire thing, start to finish. A skill is just a folder with one file inside it called skill.md. In Claude Code, you make a folder in your home directory .claude/skills/then your skill's name. Inside skill.md, you write a tiny header, a description of what the skill does and when to use it. That one line is what Claude reads to decide when to fire it. And then, below it, your actual instructions in plain English. That's it. Let me make it real. Say you want Claude to summarize your code changes. Your description says: "Use this when the user asks what changed, or wants a commit message." Then, in the body, you can even inject live data, a line that runs git diff and drops the result straight in, followed by: "summarize these changes in three bullets, and flag anything risky." Save it. Now, you either just ask, "what did I change?" and Claude loads it automatically, or you type slash, the skill's name, to run it on demand. Personal skills live in .claude/skills. Project skills go in the repo, so your whole team shares them. On claude.ai, you can upload a skill as a simple zip file. That is the same mechanism Anthropic's finance and legal teams are using, just smaller.
Once you have made one, a few things make this serious. Anthropic ships an official tool called Skill Creator, a plugin that helps you write skills and even A/B tests them, measuring whether the skill actually improves Claude's answers versus the tokens it costs. Skills also pair with MCP. Think of MCP as the wiring that connects Claude to your tools and data, and skills as the know-how for using them. Anthropic's legal team combines both. There is everything you need on the public Anthropic/skills repo on GitHub, including the real document skills you can read and copy. One honest warning, because Anthropic says it themselves: only install skills from sources you trust. A skill can run code and call tools, so treat installing one exactly like installing software. Don't paste a random skill you found online into a work machine.
So, here is your copy-this checklist. One, the finance move: take a report you rebuild on a schedule and turn it into a skill, so you review instead of rebuild. Two, the legal move: take the rules and style you repeat to everyone and turn them into a review skill that returns risk levels plus fixes. Three, start absurdly small: one folder, one skill.md file, one task you are sick of re-explaining. The teams running circles around everyone else are not more talented with AI; they just stopped re-explaining themselves and wrote it down once. Do that this week, even for one task, and you will feel the same 10 to 20 hours a week shift Anthropic's own people are talking about.
If this changed how you think about working with Claude, do me one favor: subscribe, whether you are watching on YouTube, following on Facebook, or over on Instagram. I go deep on exactly this: Claude, Claude Code, and the AI workflows that actually save time every single week, and I do the costly digging so you don't have to. Hit subscribe and the bell so the next one finds you. And if you want to go further, I have put everything together for you. Comment the word "onboard" and I'll send you a free skills starter kit, the exact skill.md templates from this video, plus the finance and legal patterns, ready to copy. The link is in the description, too. If you want the deep stuff, I have four guides linked below: my Claude Code guide, my OpenAI Codex guide, the Claude for sales playbook, and my certification prep kit, built for people who want to get genuinely dangerous with this. Comment your favorite Anthropic skill, subscribe on YouTube, Facebook, and Instagram, and I'll see you in the next one. And the description has a tip if you want to watch a little faster.
