Anthropic Skills. The Landscape for New Models and Architecture

Source: Dev.to

Introduction

Skills are modular, on‑demand data that transform general‑purpose LLMs into specialized agents. It’s not about MCP or fancy protocols – it’s about context engineering: loading the right information at the right time.

A skill is a memory, instruction, fact, or code snippet loaded on‑demand into your LLM’s context window.
A skill is also RAG + System Instructions + Domain Expertise.

Everything you can achieve with skills can technically be achieved without them: load a large number of tokens (tools, instructions, examples) into the context, use a big reasoning model with enough thinking time, and you’ll get decent results. The new skills‑based approach, however, is simpler, faster, cheaper, and more scalable.

It’s Not About MCP

I’ve been writing code for 14 years, follow AI updates closely, and consider myself a vibe‑coder. Almost a year ago I shared my thoughts about MCP from a developer’s perspective.

My core argument: MCP as an approach and programming pattern may not be the best solution.

• I don’t use MCP servers because they don’t work well for my use cases.
• Coding agents already have enough: terminal commands, file system, web‑search handle most scenarios.
• When agents must call external tools (e.g., creating a Google Calendar event), I need more robust custom code, not an MCP wrapper.
• My debugging workflow: manually add 3‑4 files + 1‑2 documentation links. This works better than any automated context retrieval.

Those files and links could be found automatically – that’s what skills promise: making agents smarter.

Context Engineering

I feel an architectural shift is coming:

• Before: Big generalized pretrained model or fine‑tuned model.
• After: Small reasoning model with new architecture to learn + Skills for task‑specific problems.

“Context engineering is the delicate art and science of filling the context window with just the right information for the next step.” — Andrej Karpathy

The key insight: it’s not about having the biggest model; it’s about having the right context at the right moment.

The Path Forward

It’s still R&D – a branch from mainstream LLM development, not a replacement. Anthropic has been working on this for 6+ months, and we are early adopters, discovering a better direction before the mainstream. My prediction: in six months everyone will rush into skills‑based agent architectures.

What we need:

• Automatic skill discovery.
• Composable skill libraries – combine skills for complex multi‑step workflows (e.g., n8n).
• Domain‑specific skill packs – pre‑built expertise for common developer tasks, such as Angular skills or GitHub Runner skills.

Conclusion

Skills embody a philosophy of modular, reusable, on‑demand expertise that transforms any LLM into a specialized agent for a specific workflow. I believe this is the right path and plan to follow its progress closely, eventually coding new things that leverage or extend skills.