For indie developers building their own AI systems, there is one clear common thread this week: Anthropic is fully committing to the principle of "build Skills, not custom agents". Instead of maintaining complex, monolithic agent frameworks, the focus is shifting to modular, reusable skills. This aligns seamlessly with the needs of developers with their own stack, local LLM gateways, and homelabs. By keeping skills small and sharply defined, they can be dynamically combined without clogging your context window. Below you will find this week's key signals and documents, translated into practical value for your setup.
Key signals from the community
The Complete Guide to Building Skills for Claude (Alternative Markdown version)
This is Anthropic's official guide on designing, testing, and distributing Skills. The document confirms an important architectural pattern: the two-layer model. In this model, the YAML frontmatter (containing metadata and activation instructions) is always loaded, while the actual body of the skill is only loaded on-demand (progressive disclosure).
Why this is relevant to your stack: If you work with a local LLM gateway like LiteLLM, this pattern helps to drastically reduce the load on your context window. Instead of sending all system instructions directly, you only load the metadata. Only when routing determines that a specific skill is needed is the full code or instruction loaded. This is directly applicable to how you structure your own SKILL.md files.
Claude Security Plugin for Claude Code (Beta)
Anthropic has released an official security plugin for Claude Code. It runs locally in your terminal and acts as a multi-agent vulnerability scanner. The plugin scans your repository, analyzes git history and data flows, and generates patch files that you can then review and apply yourself.
Why this is relevant to your stack: This is an excellent blueprint design for how to delegate complex tasks to sub-agents within a terminal environment. It demonstrates how to design a plugin that not only reads code but also integrates active feedback loops (such as generating patches for human review). You can adopt this model for your own local security and audit workflows.
Building a plugin from scratch in Claude Cowork (See also the knowledge-work-plugins repo)
A new official tutorial explains how to scaffold a plugin in Claude Cowork using natural language. You simply describe what the plugin should do, after which Claude asks targeted questions about the workflow, tools, and edge cases to then generate the code. The accompanying GitHub repository offers a wealth of concrete examples.
Why this is relevant to your stack: This makes it extremely easy to automate everyday tasks around your own agent setup or home server. Think of quickly generating a plugin that performs status checks on your storage media, or a workflow that validates backups and reports via a webhook to, for example, n8n.
Claude Cowork Guide for Power Users: 50+ Tips (Mirror available via Roger Wong)
A comprehensive collection of over fifty tested tips for working with plugins, sub-agents, and memory management in Claude. The most important advice: break your skills down into small pieces (chunking). Large, all-encompassing skills often fail because the LLM loses context. Sharp, specific descriptions ensure much better activation. Additionally, it helps to anchor specific environmental context directly within the skill files.
Why this is relevant to your stack: This aligns with the Unix philosophy: write programs that do one thing well. Apply this to your own agent setup by not building a single 'super-agent', but rather a collection of micro-skills that each handle a specific API endpoint or database query.
tonsofskills: Community Marketplace
An open-source marketplace featuring hundreds of community plugins and skills, including its own CLI package manager called ccpi. Although this is not an official Anthropic product, it clearly demonstrates how the community envisions the distribution of skills.
Why this is relevant to your stack: It is an excellent source of inspiration to see how others structure their integrations. However, keep in mind: because this is community-curated, you should always manually audit the code before granting it access to your local systems or home server.
Context and model routing
The shift toward modular skills also means that the need for smart routing is growing. If you are running locally, you don't want to call a heavy model for every simple skill. By using a central gateway, you can distribute tasks dynamically. For more background on this topic, see our guide on model routing on llmnet.nl.
What can you do with this? (Concrete actions)
- Structure your own documentation: Set up your local agent configuration according to the two-layer model. Use YAML frontmatter in your markdown files for metadata and keep the instructions in the body on-demand.
- Build a micro-plugin: Use the scaffolding method to write a simple plugin that automates a specific task on your home server, such as checking system resources or triggering an n8n workflow.
- Refactor your prompts: Go through your current prompts and break down large, monolithic instructions into smaller, reusable 'skills' with a sharp, clear description.