2026

→ A better mental model

Skills are composable, testable, and reusable units.
Agents are just orchestrators of skills.
─── Think in skills. Ship faster.

Build Skills not agents

Architecture

Architecture diagram showing four layers: AI Engineer instructs Claude (Agent), which connects to MCP Servers and Filesystem, which executes the Skills Layer of composable, testable skill units. AI Engineer Claude (Agent) MCPServer 2 MCPServer 3 Filesystem SKILLS LAYER Skillread · write · search Skillparse · transform Skillsummarise · draft Skillanalyse · report active composable callable · testable · reusable

01

AI Engineer

The human in the loop. Writes SKILL.md files, configures MCP servers, runs evals. Never writes logic inside the agent.

Writes SKILL.md Defines goals Runs evals
instructs

02

Claude
(Agent)

The orchestrator. Reads SKILL.md, decides which skill to call. Holds zero business logic — delegates everything.

MCP Server 2

Filesystem · read · write

MCP Server 3

APIs · web · search

accesses

03

Filesystem

The persistent layer. SKILL.md files live here. Outputs written here. Single source of truth for the whole system.

SKILL.md Outputs Context Logs
executes

04

Skills
Layer

Self-contained units, each defined in a SKILL.md. Composable, testable, reusable across any agent.

Active

Read · Write · Search

All filesystem & retrieval ops. Deterministic & unit-testable.

Active

Parse · Transform

JSON, Markdown, PDF → structure. Isolated bug surface.

Composable

Summarise · Draft

Structured context → written output. Upstream-agnostic.

Composable

Analyse · Report

Insights & final reports. Versioned & swappable.

The principle

Your agent's only job: decide which skill to call — and when.

What this means

  • Skills hold implementation. Agents hold intent.
  • A broken skill is a scoped bug. A broken agent is a catastrophe.
  • New capability = new SKILL.md. No agent rewrites.