Anthropic has published detailed insights into how it uses "skills" within Claude Code, a system of modular extensions that agents can discover and use to work more accurately and efficiently. The company cataloged hundreds of skills in active use internally and identified nine distinct categories that reveal how teams are structuring AI-assisted development at scale.

Skills are not simply markdown files, as some developers assume. They are folders that can include scripts, assets, data, and configuration options that allow agents to discover, explore, and manipulate resources dynamically. This distinction matters because it determines how flexible and powerful a skill can become when integrated into a development workflow.

What Are the Nine Types of Skills Anthropic Uses?

After cataloging all internal skills at Anthropic, the team noticed they cluster into nine categories. The best skills fit cleanly into one category; those that try to do too much straddle several and confuse the agent. Here is how Anthropic organizes its skill library:

• Library and SDK Skills: Explain how to correctly use a library, command-line interface (CLI), or software development kits (SDKs), including reference code snippets and common pitfalls to avoid.
• Verification Skills: Describe how to test or verify that code is working, often paired with tools like Playwright or tmux for external verification and have shown the most measurable impact on output quality.
• Data and Monitoring Skills: Connect to data and monitoring stacks, including libraries to fetch data with credentials, dashboard IDs, and instructions on common workflows.
• Workflow Automation Skills: Automate repetitive tasks into single commands, usually fairly simple but potentially dependent on other skills or model context protocols (MCPs).
• Scaffolding Skills: Generate framework boilerplates for specific functions in a codebase, especially useful when scaffolding has natural language requirements that code alone cannot cover.
• Code Quality Skills: Enforce code quality inside an organization and help review code using deterministic scripts or tools for maximum robustness.
• Deployment Skills: Help fetch, push, and deploy code inside a codebase, potentially referencing other skills to collect data.
• Debugging Skills: Take a symptom such as a Slack thread, alert, or error signature, walk through a multi-tool investigation, and produce a structured report.
• Operations Skills: Perform routine maintenance and operational procedures, some involving destructive actions that benefit from guardrails to help engineers follow best practices.

Anthropic found that verification skills have had the most measurable impact on Claude's output quality internally. The company noted that it can be worth having an engineer spend a week making verification skills excellent, using techniques like having Claude record a video of its output or enforcing programmatic assertions on state at each step.

How Should Teams Structure Skills for Maximum Effectiveness?

Anthropic shared several best practices for building and scaling skills. The company emphasized that a skill restating what Claude would do by default adds context without adding value. Instead, skills should fill genuine gaps in what an agent can accomplish alone.

One key insight is that skills benefit from clear folder structure and configuration options. Rather than treating skills as simple instruction files, teams should leverage the full capabilities of the skill system, including dynamic hooks and organized asset management. This approach allows agents to discover and manipulate resources more effectively.

The practical implication is that teams building their own AI-assisted development workflows should think about skills as a strategic investment. Skills that are well-designed and focused on a single purpose tend to work better than those attempting to cover multiple use cases. This principle applies whether teams are using Claude Code or building internal agent platforms.

Why Model Routing Is Becoming the Real Product Advantage

While Anthropic shared lessons on skills, the broader AI development landscape is shifting toward model routing as a competitive advantage. Microsoft's recent announcement of its in-house coding model, MAI-Code-1-Flash, signals that the future of AI coding tools is not about having one perfect model, but about intelligently routing different tasks to different models based on cost, latency, and quality requirements.

The key insight is that coding assistants are evolving from single-model tools into routing layers that can choose which model handles which job. Fast, lightweight models can handle autocomplete and small edits, while stronger reasoning models can tackle complex architectural planning or bug fixes. This approach reduces costs and improves performance by matching task complexity to model capability.

For teams building with AI coding tools, this means the question is no longer "which model wins?" Instead, it becomes "which jobs deserve which model, under which constraints, with which review gates?" This shift mirrors Anthropic's approach to skills: breaking down complex development work into modular pieces that can be handled by specialized tools or models.

Steps to Implement Model Routing in Your Development Workflow

Teams looking to apply these lessons can start by mapping their coding work to model classes and verification requirements. Here is a practical framework:

• Autocomplete and Small Edits: Route to fast coding models with compile error or user rejection as escalation triggers, requiring typecheck or focused test as proof.
• Test Generation: Use fast coding models as the default route, escalate on low coverage or flaky output, and require that tests actually run as proof.
• Bug Fixes: Route to strong reasoning models, escalate on multi-file causal chain complexity, and require failing tests to become green as proof.
• Migration Work: Use strong reasoning models plus repository tools, escalate on schema or API ambiguity, and require build plus targeted smoke tests as proof.
• Security Review: Route to specialized reviewers plus strong models, escalate on data access, authentication, or secrets concerns, and require human review receipt as proof.
• Documentation Updates: Use fast models as default, escalate on API or version claims, and require primary-source links as proof.

The most important column in this routing matrix is not the model itself, but the required proof. Without proof, routing becomes cost theater. A fast model that silently creates review debt is not actually cheap. A premium model that cannot show what it validated is not trustworthy. The workflow needs a receipt, not just a better answer.

Anthropic's detailed breakdown of skills and the broader industry shift toward model routing both point to the same conclusion: the future of AI-assisted development is about specialization, transparency, and measurable outcomes. Teams that invest in clear skill definitions and explicit routing policies will have better control over quality, cost, and team trust in their AI tools.