Agent Skills: Anthropic Is Pulling Another Marketing Stunt!

These past couple of days, my Xiaohongshu, Zhihu, and Bilibili feeds have been flooded with Anthropic's Agent Skills.

Watching those bloggers exclaiming "the GPT moment for Agents has arrived" and "the AI application paradigm has been completely overturned" is honestly hard to watch. AI is indeed a hot topic now, and it's easy to make money off the traffic, so a bunch of traffic-chasers come out posting this kind of filler content. Blowing a "Prompt organization scheme" from engineering into some kind of black technology—this kind of information noise is of no value to actual developers.

1. Isn't this just the default play for everyone?

When building complex Agent systems, as features increase, you inevitably face a problem: how to handle multiple different Prompts?

Let's look at how we used to do it:

• Fixed Workflow stage: The most primitive approach, hardcoding the flow. Step one loads Prompt A, passes the result to Step two's Prompt B. All logic is hardcoded in code—stable but extremely inflexible; adding one feature means changing a ton of code.
• Intent classification (Router) stage: Later, people got smarter and put an intent classifier at the entry point. The Agent first goes through a classifier to determine whether the user wants polishing, translation, or data lookup. After classification, it loads the corresponding rules from the Prompt library. This is actually the core logic of most autonomous Agents or RAG systems today.

So what did Anthropic's Agent Skills do? It merely standardized this set of logic that everyone tacitly agreed on and was already running by default.