Agent Development Kit with Claude: From Chatbot to Operational Platform
Imagine you hire an exceptionally talented employee: quick to learn, writes good code, creates smooth documentation, and analyzes data keenly. But every morning when they arrive at the office, this person forgets everything from the previous day — they don’t remember the project they are working on, don’t recall the naming conventions, and don’t remember how you like things presented. Every day you have to onboard them from scratch.
That is exactly the feeling when using a bare AthI assistant for real work. The model is very intelligent, but each conversation is a blank slate. For AI to truly integrate into business operations, it needs more than intelligence: it needs memory, expertise, tools, authority, control mechanisms, delegation capabilities, and reusability.
This is the introductory article of the "Agent Development Kit with Claude" series — a roadmap of the mechanisms that transform Claude from a disjointed Q&A chatbot into an operational work platform. By the end of this series, you will understand the six layers necessary to design an "AI employee" for your department.
A note on terminology before proceeding. "Agent Development Kit (ADK)" is actually the name of a framework from Google, not from Anthropic. In this series, "Agent Development Kit" is just an editorial label for clarity. From Anthropic's perspective, the two official names you should remember are Claude Code (the product) and Claude Agent SDK (the SDK for programming agents). We will always refer back to the correct terminology.
Four Confusing Products
Before discussing the layers, it is important to distinguish between four commonly confused items:
- Claude Code — an agent running in the terminal/IDE (and also desktop, web): reads codebases, edits files, executes commands. This is where most of the mechanisms in this series "live."
- Claude Agent SDK — an SDK for you to build your own agent on the same framework (harness) as Claude Code. It was originally named "Claude Code SDK" and was renamed due to a broader vision: to provide the agent with "a computer" (file system + terminal + tools) to perform tasks beyond programming, such as finance, personal assistance, or research (Anthropic Engineering).
- Claude apps (claude.ai) — chat applications and Projects for end-users.
- Claude Cowork — an agent running on desktop for multi-step knowledge work, using the same agent architecture but not requiring a terminal (Anthropic).
Common point: the same "brain" Claude, but how you navigate and equip it determines whether it is a chatbot or an operational employee.
Four Pillars Transforming Claude into an Operational Platform
Why do some teams successfully implement AI while others only stop at "casual Q&A"? The difference lies in four pillars:
- Context (context/memory) — Claude needs to know what the business does, how the project is structured, and what the conventions are. This is the CLAUDE.md layer.
- Skills (expertise) — packages the how-to of a specific task (writing proposals, building landing pages, reading contracts) into modules that can be loaded on demand. This is the Agent Skills layer.
- Tools (tools) — connects Claude with external data and systems (GitHub, databases, CRM, Google Drive...). This is the MCP layer.
- Governance (control) — ensures the agent acts safely, follows processes, and prevents dangerous actions. This is the Hooks layer (along with permissions and managed settings).
Additionally, there are two organizational pillars: assignment (dividing tasks among multiple child agents — Subagents) and distribution (packaging capabilities to share with the entire team — Plugins).
"Steering" Claude: 7 mechanisms, and one important truth
Anthropic describes a set of mechanisms to steer Claude Code, consisting of seven elements: CLAUDE.md, rules, skills, subagents, hooks, output styles, and append system prompt (Anthropic — Steering Claude Code). This sequence focuses on the layers with the most practical weight for businesses.
There are two things not included in the "steering" list but are extremely important, and the sequence still teaches them thoroughly:
- Plugins are not the "eighth steering mechanism." A plugin is a way to package and distribute: bundling skills, subagents, hooks, MCP server, and slash commands into a single unit installed with a command (Anthropic — Claude Code plugins).
- MCP (Model Context Protocol) is an open standard to connect Claude with external tools and data — it adds capabilities, not steering behavior (Anthropic — MCP).
And here is the most important truth of the entire sequence, which should be etched in your mind:
"Guidance" is different from "guardrail." CLAUDE.md, rules, skills, output styles are context — Claude will try to adhere to them but does not guarantee absolute compliance. To ensure an action definitely occurs (or is definitely blocked), you need a deterministic mechanism: that is hooks and permissions/managed settings (Anthropic — Steering Claude Code).
In other words: writing "Absolutely do not delete the database" in CLAUDE.md is misusing the tool. To truly block it, use a hook. Understanding this boundary will help you avoid a lot of disappointment during implementation.
Six layers this sequence will teach
Each subsequent article will dissect one layer, deep enough for you to apply it immediately.
Layer 1 — CLAUDE.md (memory): the agent's "constitution"
CLAUDE.md is the Markdown file that Claude reads at the beginning of each session to have sustainable context: what the project does, conventions, build commands, architecture (Claude Docs — Memory). There is a hierarchy from the organization as a whole down to the individual, with commands /init and /memory, and even an automatic memory mechanism. Why it matters: without this layer, each session is a re-onboarding. → Article 1: CLAUDE.md
Class 2 — Agent Skills (expertise): on-demand capabilities
A skill is a directory containing the file SKILL.md, packaging a process or expertise. Claude only loads the name + description of the skill initially (very low token cost), and only loads the entire thing when the task actually requires it — the principle of "progressive disclosure" (Anthropic — Skills explained). Why it matters: instead of cramming all knowledge into a huge prompt, you package capabilities into modules that activate when needed. → Article 2: Agent Skills
Class 3 — Hooks (control): deterministic guardrails
A hook is a command/script that runs automatically and deterministically at fixed points in the agent's lifecycle — for example, automatically formatting code after a file is edited, or blocking a dangerous command before it runs. The model does not decide whether the hook runs or not. Why it matters: this is the "real guardrail" for AI in business. → Article 3: Hooks
Class 4 — Subagents (delegation): AI assistant teams
A subagent is a separate instance of Claude, with its own context window, tools, permissions, and tasks (Anthropic). The main agent coordinates; subagents handle the deep work and only return results — keeping the main context tidy. Why it matters: one agent should not take on everything, just like a company needs multiple roles. → Article 4: Subagents
Class 5 — Plugins & Marketplace (distribution): turning capabilities into assets
A plugin bundles skills + subagents + hooks + MCP + commands into a single installable package, which can be distributed through a marketplace (Claude Docs — Plugins reference). Why it matters: AI capabilities are no longer confined to one individual's mind but become operational assets for the entire organization. → Article 5: Plugins & Marketplace
Class 6 — MCP Servers (connection): plugging Claude into the outside world
MCP is an open standard that enables Claude to communicate with external tools and data without having to write custom integrations for each service (Anthropic — MCP). Why it matters: an agent that cannot connect to real data will always remain theoretical. → Article 6: MCP Servers
Quick comparison table of the six classes
| Class | What problem does it solve | When is it loaded into context | Who/what controls it |
|---|---|---|---|
| CLAUDE.md | Context & sustainable conventions | At the start of each session (always present) | You write it; Claude tries to adhere |
| Skills | Expertise/process on demand | Only name + description initially; full when needed | Claude self-identifies when relevant |
| Hooks | Guardrails & automation | Triggered by lifecycle events | Deterministic — always runs, not model-controlled |
| Subagents | Delegation, context isolation | Runs in its own context | Main agent delegates per description |
| Plugins | Packaging & distribution | When installing/enabling the plugin | Installer; managed by the organization |
| MCP | Connecting external tools/data | When declaring server + calling tool | Open standard; has approval & permissions |
How to read this string
Six layers are arranged in a natural flow: CLAUDE.md → Skills → Hooks → Subagents → Plugins, with MCP as the connecting element throughout. You can read sequentially from Article 1 to build a solid foundation, or jump straight to the layer you need. Each article includes real-life examples (accounting, marketing, programming) and citations from the original documents for you to verify and delve deeper.
Another point: in-depth Vietnamese documentation on these layers (especially Hooks and Plugins) is currently quite sparse — some good content is available in the Claude Code series from AZDIGI. This series aims to become a comprehensive and accurate Vietnamese resource, always anchored to the original documentation from Anthropic.
In Article 1, we start from the foundational layer: CLAUDE.md — how Claude "remembers" your project, organizes memory hierarchically, and best practices that make this file truly effective instead of unnecessarily bloated.
References
- Anthropic — Steering Claude Code (framework of 7 navigation mechanisms): https://claude.com/blog/steering-claude-code-skills-hooks-rules-subagents-and-more
- Anthropic — Skills explained: https://claude.com/blog/skills-explained
- Anthropic Engineering — Building agents with the Claude Agent SDK: https://www.anthropic.com/engineering/building-agents-with-the-claude-agent-sdk
- Claude Docs — Memory (CLAUDE.md): https://docs.claude.com/en/docs/claude-code/memory
- Anthropic — Customize Claude Code with plugins: https://www.anthropic.com/news/claude-code-plugins
- Claude Docs — Plugins reference: https://code.claude.com/docs/en/plugins-reference
- Anthropic — Introducing the Model Context Protocol: https://www.anthropic.com/news/model-context-protocol
- Anthropic — Claude Cowork: https://www.anthropic.com/product/claude-cowork
- AZDIGI — Vietnamese Claude Code series: https://azdigi.com/blog
Article 0/7 in the "Agent Development Toolkit with Claude" series. The content has been rephrased from the original documentation of Anthropic to comply with copyright constraints. Next → Article 1: CLAUDE.md — Memory layer.