How to Use Agent Skills

March 18, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
March 18, 2026

Last call for Agent Madness!

Join our March Madness style bracket competition for agent builders. Show off the coolest agent you’ve been working on and win the chance to have your agent featured on the show. Go to agentmadness.ai to enter.

HEADLINES

Claude Cowork Dispatch: Control Your Agent from Your Phone

On the surface, Dispatch is a simple feature: you can now kick off a Claude Cowork session on your desktop and manage it from your phone. But based on the reaction — 3 million views on the announcement tweet, 9,000 bookmarks — this one clearly landed. Cowork developer Felix Rieseberg described it as "like having a walkie talkie for communicating with Claude." Ethan Mollick tested it and wrote that Dispatch covers 90% of what he was trying to use OpenClaw for, but "feels far less likely to upload my entire drive to a malware site." As Pawel Huryn put it: Code, Cowork, Web, and now Dispatch are all converging toward the same thing — a persistent AI layer that follows you across devices and contexts. That's exactly right: this ongoing qualifiction of everything is really just form factor adjustment as everyone figures out how people actually want to interact with agents.

China's OpenClaw Problem

OpenClaw is a cultural phenomenon in China — videos of people lining up for installation services, some local governments offering subsidies for entrepreneurs building on it — but the government is now growing concerned. Regulators have warned state agencies and state-owned enterprises not to install it, and last week issued a formal list of six dos and don'ts. Adoption is so pervasive that the Hong Kong Monetary Authority had to issue a statement that they had no plans to deploy it on internal IT systems. Meanwhile, China's AI anxiety spike is notable: while optimism about AI remains higher there than in the US, the term "AI anxiety" surged on WeChat in February, peaking in mid-March. Tony Peng of Recode China explains why: for most ordinary people, AI still meant chatbots — then suddenly media reports were claiming OpenClaw can handle a wide range of tasks autonomously, and the gap between what people knew and what they were being told deepened the sense of being left behind. Separately, Chinese authorities are also taking a second look at Meta's acquisition of Manus, with officials apparently trying to prevent Manus executives from departing China for Singapore ahead of trade talks.

NVIDIA Restarting China H200 Production

Jensen Huang confirmed at a press conference Tuesday that NVIDIA is restarting manufacturing for H200 chips destined for China, with purchase orders already in hand. The process of getting export approval has been on-again, off-again since Trump floated the idea in December. Huang's read on how it got resolved: "President Trump's intention is that the United States should have a leadership position in access to Nvidia's best technology. However, he would like us to compete worldwide and not concede those markets unnecessarily." Reuters confirmed Chinese authorities have approved purchases for multiple companies. Demand is reportedly staggering, which helps explain where that $1 trillion revenue forecast is coming from.

Amazon CEO: AI Could Double AWS Revenue to $600B by 2036

Amazon CEO Andy Jassy told staff at a recent all-hands that AI could boost annual AWS sales to $600B — double his prior estimate. AWS booked $128B in 2025 at 19% growth, so at that rate the number isn't extravagant: it's roughly 17% annual growth sustained for a decade. Patrick Moorhead called it "the clearest signal yet that hyperscale cloud is entering a second growth phase that dwarfs the first." The more interesting framing: if this plays out, Amazon becomes one of the biggest beneficiaries of the entire AI buildout without having had to build the models themselves.

MAIN STORY

A Primer on Agent Skills

Skills are one of the most important and underappreciated concepts in the current agentic era — and a post from Tariq on the Claude Code team at Anthropic, "Lessons From Building Claude Code: How We Use Skills," gave the most concrete and practical treatment of the topic yet. This episode uses it as a jumping-off point to explain what skills actually are, how to think about them at different levels of technical sophistication, and where the format is heading across the entire agent ecosystem.

SKILLING UP

What Skills Are
"Write once, use everywhere."
Skills were born out of a specific problem: as AI coding agents got more capable, system prompts kept ballooning. Every new capability meant more instructions and examples crammed into a single context window — which caused performance degradation, slower and more expensive agents, and less reliability. The insight that drove skills is that agents don't need all their knowledge all the time. They need to load the right knowledge at the right moment. A skill is a folder anchored by a SKILL.md file with a name and short description. The agent sees those descriptions while working, pulls up the skill if it seems relevant, and gets progressive disclosure of more detailed context only as needed — from description, to body, to any linked scripts or reference files. The analogy: a well-organized manual that starts with a table of contents, then specific chapters, then a detailed appendix. Skills launched on October 16th with Anthropic's blog post "Equipping Agents for the Real World With Agent Skills" and were almost immediately adopted beyond just Claude — OpenAI added skills support to ChatGPT and Codex, GitHub Copilot adopted the standard, and a community skills hub called ClawHub now hosts around 28,000 skills.

Anthropic's Skill Taxonomy
"Verification skills alone can be worth an engineer's week."
When the Claude Code team audited the 28,000 skills on ClawHub, they found that most fell into nine categories: library and API reference, product verification, data and analysis, business automation, scaffolding and templates, code quality and review, CI/CD and deployment, incident runbooks, and infrastructure ops. Two categories worth highlighting for this audience: business process and team automation (skills that automate repetitive workflows into one command — the example is a weekly recap skill that merges PRs, closes tickets, and formats a recap post) and code quality and review (adversarial review skills that spawn a fresh-eye subagent to critique until findings degrade into nitpicks; code style skills that enforce patterns Claude doesn't do well by default). The highest ROI category, Tariq argues, is verification skills — skills that describe exactly how to test whether Claude's output is correct, including techniques like having Claude record a video of its output so you can see precisely what was tested.

The Skill Creator Update
"Evals, benchmarks, and keeping skills working as models evolve."
Alongside Tariq's post, Anthropic quietly updated their Skill Creator tool — which solopreneur Ali Lemon flagged as a bigger deal than it looks. The update addresses three real problems: (1) no way to measure how well your skills were actually performing — now you can run evals against multiple prompts and get a score; (2) skills breaking silently when models update — now you can run A/B tests comparing your skill against raw Claude; (3) Claude not using your skill because the description is too vague or too specific — the Skill Creator now rewrites descriptions automatically to trigger at the right time. Anthropic tested this on their own skills and saw better triggering in five out of six cases. The Skill Creator also introduces a useful two-category framework: capability uplift skills (helping Claude do something it can't do consistently, like specific document formats) versus encoded preference skills (sequencing things Claude could already do according to your team's workflow). Capability uplift skills may become less necessary as models improve; encoded preference skills are more durable but only as valuable as their fidelity to your actual process.

Best Practices from the Claude Code Team
"The highest signal content in any skill is the gotcha section."
Tariq's top tips: Don't state the obvious — if a skill is primarily about knowledge, focus on information that pushes Claude out of its normal way of thinking. The front-end design skill at Anthropic was built by iterating with customers to improve Claude's design taste, specifically by calling out classic patterns to avoid (the Inter font, purple gradients). Build a gotcha section — articulate common failure points and update it over time so the skill becomes a living document that gets smarter with every mistake. Think of the entire file system as context engineering, not just a single markdown file. And give Claude the information it needs without railroading it — flexibility to adapt matters.

Skills Across the Whole Stack
"From consumer tools up to complex multi-agent teams."
The concept of skills as reusable capabilities is converging across the entire AI stack simultaneously. For advanced agent builders, skills are a modular architecture for agent capabilities. For individual power users — probably most listeners — skills are reusable prompts with superpowers: the difference is that a skill can include actual code templates, reference data, and examples, not just instructions. You figure out how to get the agent to do something well once, package it, and it works reliably every time. For mainstream users not yet in Claude Code or Codex, the design pattern still holds: this week Notion announced custom skills for Notion AI, with the framing "write a prompt and you'll use it once; write a skill and you'll use it forever." The underlying shift is the same at every level — AI is less and less a one-off conversation, and more and more a library of reliable, repeatable capabilities.