Cal Rueb

I highly recommend AI engineers take a close look at our Claude Skills release today. Under the hood is a reusable pattern for anyone building agents.

Here is my simple pitch. If you're building on LLMs you know well that there is a tipping point where if you fill up your prompts with too many instructions it starts to water down the model's ability to closely follow all of them. Your agent might only follow important instructions some of the time, or ignore them completely.

I have seen this first hand with teams building agents to do customer support. It turns out you can't take the 30 page standard operating procedure (SOP) you give the support team and just dump that into a prompt and expect the agent to follow everything perfectly. There are too many instructions and it overwhelms the model.

Additionally, most instructions are only relevant some of the time. Only once I know a customer wants to initiate a return do I need to know all the specifics of how to do that.

The skills pattern allows you to provide just-in-time instructions to your agent. The idea is you write conditional instructions in separate files, give your agent a way to search and read these files, and let the model pull these instructions into context only when the agent believes they will be relevant. This protects the context window, and prevents your instruction following from getting watered down.

I believe this pattern is highly relevant for teams building any sort of agent with large sets of instructions you are providing at run time.

Our engineering blog post goes into a lot more detail on how this works under the hood (https://lnkd.in/ga-gjRwE).

I am proud to work at Anthropic. Our Applied AI team remains focused on helping startups build world-changing products on Claude.

https://lnkd.in/gUmyEVJZ

One regular complaint we get from teams building coding tools and app builders is that Claude does not produce visually interesting frontends.

Claude will create webpages and apps that visually are bland and repetitive - typically overusing the color purple and a heavy use of emojis.

My teammate Prithvi R. dug into this feedback and quickly found that yes this is a real issue - HOWEVER with some thoughtful prompting you can get Claude to produce much richer outputs.

He wrote a cookbook capturing those learnings and includes a set of instructions to put in your prompt that produces much better results. Check it out here: https://lnkd.in/gbwnHAM2

As a quick test I gave Claude Code running Opus 4.1 the task "'Create an HTML page that celebrates "Claude is good at design".' with and without appending Prithvi's prompt in my user message. Pictured are the results side by side!

An important AI engineering lesson I often repeat - the models you are building on top of right now are the least intelligent, slowest, most expensive they are ever going to be.

This trend continues today with Haiku 4.5 - a model who's agentic capabilities go toe-to-toe with Sonnet 4 in a sleeker, faster, cheaper package.

Happy building!

When building long running agents you are going to run into challenges like "how do I keep my agent running coherently after I fill up the context window" or "how can I give multiple sessions of my agent a persistent memory to work off of".

There are several techniques for dealing with these challenges, and often these techniques overlap and it's not clear what to use when.

Isabella He wrote a detailed cookbook on three of these techniques: tool clearing, compaction, and file-system based memory. In it she compares all three against the same long-running task so you can see exactly when each one kicks in, what it costs, and how they layer together.