Andrew Ng’s Agent Skills: From Basics to Must‑Know Essentials

Agent Skill definition

Agent Skill = an organized folder containing instructions, scripts, assets and resources that enable an agent to execute a specific task precisely.

analyzing-marketing-campaign/
├── SKILL.md          ← core instruction file
└── references/
    └── budget_reallocation_rules.md  ← reference resources

Motivation: from siloed agents to a universal agent

Earlier systems built a separate agent for each domain (Research, Coding, Marketing, Finance). Each agent operated in isolation, leading to high maintenance cost—a classic anti‑pattern.

The new paradigm introduces a single generic agent combined with a skill library. The generic agent lacks domain‑specific knowledge (brand guidelines, legal review steps, marketing metrics); skills supply that missing expertise.

What Skills add – three capability dimensions

Domain expertise : provides industry‑specific rules such as brand guidelines, legal processes, and data‑analysis methodology.

Repeatable workflow : encodes standardized procedures for recurring tasks (weekly marketing recap, client‑call preparation, quarterly business review).

New capabilities : enables actions the base agent cannot perform, e.g., generating PPT decks, creating Excel/PDF files, building MCP servers.

Pain points without Skills

Re‑describing instructions for each run – error‑prone and time‑wasting.

Re‑packaging reference files – scattered context makes collaboration difficult.

Inconsistent output – results are unreliable and cannot be scaled.

Core properties of Skills

Portable – write once, run everywhere

The same Skill can run seamlessly on: Claude Code (CLI agent) Claude.ai (web chat) Claude Agent SDK (self‑built agent) Claude API (direct calls) Other AI tools such as Cursor, Trae, Codex, etc.

Skills are an open standard; a Skill authored today can be executed on any compatible platform tomorrow.

Composable – Lego‑style workflow construction

Multiple Skills can be chained to implement an end‑to‑end business process. Example: generating a marketing analysis report.

BigQuery Skill          → fetch marketing DB schema
↓
Marketing Analysis Skill → analyze multi‑channel data
↓
Company Brand Skill      → retrieve brand guidelines
↓
PowerPoint Skill        → auto‑generate presentation

Four independent Skills replace the effort of an entire team.

Progressive disclosure loading strategy

Hundreds of Skills may exist, but LLM context windows are limited. The solution is to load Skill information in three layers only when needed.

Metadata – "what I can do", always loaded (analogy: employee résumé).

Command – "how I should act", loaded on trigger (analogy: operations manual).

Resource – specific rules or assets, loaded on demand (analogy: appendix).

Takeaways for engineers

Skills constitute knowledge engineering; industrial‑scale AI systems require structured knowledge management rather than ad‑hoc prompt piles.

Open‑standard Skills create ecosystem benefits comparable to component libraries or API specifications.

Progressive disclosure enables a single agent to handle unlimited complexity within a bounded context.