Claude can now use Skills. Skills are packaged instructions that teach Claude your way of working.

Skills work across Claude Code, and the API. Claude identifies relevant Skills and applies them automatically. They can also stack together for complex workflows.

Enable Skills and build your own in Settings > Capabilities > Skills. Available now for all paid plans. Read more:

if claude messes it up, im replying "skill issue" be ready for that lol

"Skill issue" now got a new meaning 👀

Claude:

Even if Claude is incredibly capable, when I give it a 10-page PDF and say “look at this,” the conversation context runs out after just a few follow-up questions. No matter how good it is, if you can’t actually use it properly, it’s pointless.

my .md files right now

Cool feature. Here is how Skills differ from / command if anyone is wondering. I already used / commands this way, but I think it's nice to have "skills" natively supported and understood by Claude and to use when needed.

Skill is such a great idea that it makes me wonder: why hadn't anyone thought of this earlier? I can think of SKILL in 2 ways: 1. It's a 'soft workflow'. Workflows are rigid, but if you define your workflow in English in skill .md, the agent's intelligence can make it more dynamic and flexible. 2. It streamlines the system prompt. In the old days (2 months ago?), the system prompt is bloated by the many tool definitions and examples. Now you can just include the meta-info of a skill in system prompt, when the skill is needed, the agent can look up skill .md to learn to use the tool. And the skill definition can be further divided into components! This can save a lot of context space. There will be a lot of agent-skill repos appearing on github, just like when MCP was introduced.

Next update: Claude learns how to gaslight humans professionally

how does it differ from

this looks interesting. not because of what they do now, but because they're composable and portable. you can stack multiple skills, they work across apps/api/code, and they auto-load when relevant. this is the Unix philosophy applied to AI agents. whoever builds the best skill library owns the ecosystem

So it was having a skill issue previously?

For those wondering how this works in Claude Code, I made this video

Very exciting! A lot of Claude Code plugin authors are adding these reusable skills to their plugins. You can now browse and filter plugins by skills on

nice claude skills available on claude code too!

This is genuinely one of the most interesting updates. 😳 On the surface people might miss how big of a deal this is. I recently bought a recording setup and I’m excited to show people how powerful this can be for creative use cases. (Will need to experiment first to confirm it is as good as it appears)

The engineer blog link is giving me a 404

i'm so tired

How is this different from Projects + Artifacts or is this just re-packaging a process to give it a fancy name. Or am I missing something?

how does it differ from any other md file that's used? can't you just generalise it all as rules?

Yeah but then u retain our data for 5 years!!!?

Subagents , slash commands , files now skills

I like these a lot. I would argue a lot of agents are really just "skills".

Guys, wtf is this?

All of these upgrades are practically unusable if your context window maxes out in a few rounds of back and forth discussion

Skills. zip because apparently intelligence now comes in compressed files

When ChatGPT's memory feature launched, I wanted to use it to define commands programmatically or correct small repeatable mistakes by the model Skills look like they'll provide that and go further. They're also standardized and sharable. The Ants have earned their mandate

great. When sonnet-[1m]??

Claude doesnt follow his in the first place, why would he follow this? Opus 4.1 ClaudeCode screenshot provided.

how do i make a video like this with the screen-sharing and the highlighting of different parts?

Skills could make Claude fit into existing workflows by encapsulating repeatable instructions. Recommend versioning, clear inputs and human review hooks to keep behavior predictable. A pilot with task level metrics will prove ROI quickly. 🤖

There's a problem with the blog post, it links to an internal anthropic login page

if it messes up, I'll call it "Skill issue"

Haven't used Opus for the last week. I change to Opus, type a 12 line prompt and get a warning "Approaching Opus usage limit". What am I paying $100USD/month for?