AGENTS & ORCHESTRATION

The Skills Paradigm

One agent, infinite expertise

In 30 Seconds

The old way: Build a separate AI agent for each domain. A finance agent, a legal agent, a research agent – each with its own scaffolding, its own tools, its own maintenance burden.

The new way: One general-purpose agent + a library of skills. The agent loads relevant expertise on demand. New domain? Add a skill file. That's the paradigm shift.

Anthropic's realisation: “After we built Claude Code, we realized that Claude Code is actually a general purpose agent.” Skills let any organisation extend AI capability across domains – without building new agents.

Why This Matters

For Growing Organisations

You don't need an AI team to build agents. Skills are simple files – instructions, examples, references – that anyone can create.

One person's expertise becomes organisation-wide capability. Knowledge compounds instead of staying locked in individual heads.

For Technical Teams

No more maintaining separate codebases for each domain agent. The runtime is the same – only the loaded context changes.

Version skills in Git. Share them across projects. The infrastructure stays constant while capability grows.

The democratisation: A few companies build AI models. A few more build agent runtimes. But anyone can build skills. This is the layer where domain expertise becomes AI capability.

Skills vs Agents: The Key Distinction

Understanding when to use each

Skills

Procedural knowledge loaded into the main agent's context. The agent reads the skill and knows how to do something new.

What they are

Markdown/text files containing instructions, examples, best practices, and reference material

When to use

“How do I...?” questions. Domain expertise. Standard procedures. Best practices.

Example

Financial reporting skill: templates, regulations, common pitfalls, quality checks

Subagents

Separate agent instances spawned to work autonomously. They run in parallel, with their own context, and return results.

What they are

New agent processes with their own context window, spawned by the main agent

When to use

“Do X for me” tasks. Parallel work. Isolated execution. Long-running operations.

Example

Code review subagent: analyses code independently, returns findings when done

The simple test: Does the main agent need to know how to do something? Load a skill. Does something need torun independently? Spawn a subagent.

Anatomy of a Skill

Skills are deliberately simple. Here's what goes inside.

Essential Components

Description

What this skill does and when to use it. The agent reads this to decide whether to load the skill.

Instructions

Step-by-step guidance. What to do, in what order, with what considerations. The procedural knowledge itself.

Examples

Show, don't just tell. Good examples of inputs and outputs. What “done well” looks like.

Optional Enhancements

Reference Material

Domain knowledge the agent might need. Definitions, frameworks, regulations, or standards relevant to the skill.

Common Pitfalls

What to avoid. Known failure modes, edge cases, and mistakes to watch for.

Quality Checks

How to verify the work is good. Checklists, validation criteria, or review steps.

The format is intentionally low-tech. Plain text or Markdown. Version in Git. Edit with any tool. No special infrastructure required – just well-organised knowledge.

Types of Skills

Three categories, each serving different purposes

Foundational Skills

New capabilities the agent didn't have before. Domain expertise packaged for AI use.

Document Creation

Reports, proposals, analyses with domain-specific structure

Data Analysis

Statistical methods, visualisation patterns, interpretation frameworks

Research

Literature review, synthesis, citation patterns, source evaluation

Technical Writing

Documentation standards, API references, user guides

Platform Skills

Help agents work better with specific tools and platforms. Usually built by platform providers.

Browser Automation

Web navigation, form filling, data extraction from pages

IDE Integration

Code editing patterns, project navigation, build systems

API Orchestration

Multi-step API workflows, error handling, rate limiting

Database Operations

Query patterns, schema navigation, data migration

Enterprise Skills

Organisation-specific knowledge. Your unique processes, standards, and institutional knowledge.

Internal Processes

How things get done here. Approval flows, standard procedures

Coding Standards

Your team's conventions, patterns, and review criteria

Compliance Rules

Regulatory requirements, audit trails, documentation standards

Domain Knowledge

Industry-specific expertise unique to your organisation

How Skill Loading Works

The agent decides what expertise it needs, when it needs it

1

Task Arrives

User asks the agent to do something. The agent assesses what kind of expertise might be needed.

2

Skill Discovery

The agent reviews available skills. Each skill has a description explaining what it does and when to use it. The agent matches task to skill.

3

Context Loading

The relevant skill is loaded into the agent's context. The agent now has access to instructions, examples, and reference material.

4

Execution

The agent completes the task using the loaded expertise. The skill provides the “how”; the agent provides the reasoning and execution.

The elegance: The agent is always the same. What changes is what it knows. New capability = new skill file.

Building Your Skills Library

Start simple, expand based on need

Start Here

Identify Repeated Work

What do you explain repeatedly? What processes follow consistent patterns? These are your first skill candidates.

Capture Expert Knowledge

What does your best person know that others don't? Document it as a skill so AI can apply it consistently.

Start With One Skill

Don't build a library upfront. Create one skill, use it, refine it. Then build the next based on real need.

Grow Over Time

Let AI Create Skills

Agents can write skills too. After completing a task, ask them to document it as a skill for future use.

Version and Share

Skills in Git means version history, collaboration, and sharing across teams or projects.

Refine Based on Use

Skills improve through use. What worked? What was missing? Update the skill file after each significant use.

The compound effect: Each skill created becomes permanent organisational capability. Knowledge stops walking out the door when people leave.

Common Mistakes to Avoid

Building Too Complex

Skills should be simple text files. If you're building elaborate infrastructure, you're overengineering. Start with a markdown file.

Designing Before Using

Don't plan a 50-skill architecture upfront. Build one skill, use it, learn what works. Let the library emerge from real needs.

Forgetting Maintenance

Skills need updating as knowledge evolves. Build review rhythms into your process. Stale skills produce stale outputs.

Hiding Skills in Silos

Skills compound when shared. One team's expertise should become organisation-wide capability. Make skills discoverable.

How We Help

Design, build, and evolve your skills architecture

Skills Discovery

Identify where skills will have most impact. Map your expertise, find the patterns, prioritise what to build first.

Skills Development

Help create your first skills. Extract knowledge from experts. Build the templates and patterns for your organisation.

Architecture Design

How should skills be organised? How do they interact with your agent setup? Design the system that scales.

Build Expertise That Compounds

Skills turn individual expertise into organisational capability. One file at a time, your AI becomes genuinely expert in your domain.

No commitment, no pitch. Just a conversation about where skills could extend your AI capability.

Disclaimer: This content is for general educational and informational purposes only. Skills architecture concepts reflect current best practices and Anthropic's published guidance. For specific guidance on AI implementation in your organisation, please consult appropriately qualified professionals.