Getting the Most Out of Claude in 2026: Skills, Hooks, MCP and the Agent SDK
If you set up Claude Code a year ago and haven’t touched the config since, you’re using maybe a third of what it can do now. The developer surface around Claude has expanded fast: skills, hooks, subagents, plugins, MCP servers, a full Agent SDK and a new model lineup with adaptive thinking. Each one solves a different problem, and knowing which tool fits which job is where the productivity gains actually come from.
We use most of this daily at Studio Cavan. Here’s the practical breakdown.
The Model Lineup First
Everything else sits on top of the models, so start there. The current lineup as of mid-2026:
- Claude Fable 5 - the newest and most capable model, first in the Claude 5 family and a tier above Opus. Built for the hardest reasoning and long-horizon agentic work, where a single request can run for many minutes while the model gathers context, builds and verifies its own output. Thinking is always on. Priced above Opus at $10/$50 per million tokens, so reserve it for work that earns it.
- Claude Opus 4.8 - the flagship of the Opus tier and the right ceiling for most teams. Drives overnight refactors and multi-hour autonomous coding sessions. 1M token context window, 128K output.
- Claude Sonnet 5 - near-Opus quality on coding at a third of the price. For most day-to-day development work this is the right default.
- Claude Haiku 4.5 - fast and cheap. Use it for classification, routing and subagent grunt work where you don’t need deep reasoning.
Fable 5 also ships alongside Claude Mythos 5 - the same underlying model without certain dual-use safety measures, available only to approved organizations. For everyone else Fable 5 is the top of the stack. Anthropic’s announcement post covers the distinction.
Two API changes matter if you’re building on these directly. First, manual thinking budgets are gone - the newer models use adaptive thinking, where the model decides how much to reason per request (on Fable 5 you can’t turn thinking off at all). Second, there’s an effort parameter (low through max) that controls the depth-versus-cost tradeoff. For agentic coding, xhigh is the sweet spot and it’s what Claude Code uses by default.
The full model docs live at platform.claude.com/docs.
Skills: Stop Pasting the Same Instructions
A skill is a folder with a SKILL.md file in it. That’s the whole format. Drop it in .claude/skills/deploy/SKILL.md and you get a /deploy command, plus Claude will load it automatically when the task is relevant.
The signal that you need a skill: you keep pasting the same checklist into chat, or a section of your CLAUDE.md has grown from facts into a procedure. Facts belong in CLAUDE.md because it loads into every session. Procedures belong in skills because the body only loads when it’s used - long reference material costs almost nothing until you need it. That distinction is called progressive disclosure and it’s the reason skills scale where a bloated CLAUDE.md doesn’t.
Skills we actually run: a release checklist that verifies the build, checks for uncommitted files and deploys to Cloudflare. A blog post skill that enforces our style rules (this site bans em dashes and Oxford commas - the skill knows that). A spec-writing skill with our template baked in.
Skills follow the Agent Skills open standard, so they work across tools, and custom slash commands have been merged into the same system. One format, two invocation paths: you call it with /name or Claude picks it up on its own.
Docs: code.claude.com/docs/en/skills
Hooks: Rules That Can’t Be Ignored
Skills and prompts rely on the model choosing to follow them. Hooks don’t. A hook is a shell command that runs deterministically at a lifecycle event - before a tool call, after a file edit, when Claude finishes a task. The model can’t skip it because the model isn’t in the loop.
This is the right tool any time you catch yourself writing “ALWAYS” or “NEVER” in a prompt. Some patterns that earn their keep:
- PostToolUse - run prettier or your formatter after every file edit. No more asking Claude to remember formatting.
- PreToolUse - block writes outside the repo, or reject any shell command that touches production credentials.
- Stop - run the test suite when Claude says it’s done. If tests fail, Claude sees the output and keeps working.
The mental model: prompts are requests, hooks are enforcement. If a rule breaking would cost you real time or money, it belongs in a hook, not a prompt.
Docs: code.claude.com/docs/en/hooks-guide
Subagents: Parallel Work Without Context Pollution
A subagent is an isolated Claude instance with its own context window, its own tool permissions and optionally its own model. The main session delegates a task, the subagent does the work in its own context and only the result comes back.
Two reasons this matters. First, parallelism - fan out ten subagents to review ten files simultaneously instead of reading them serially. Second, context hygiene - a subagent can churn through 200K tokens of search results and hand back a three-paragraph summary. Your main session never sees the noise.
The practical pattern: keep your main session on the expensive model for judgment and orchestration, push mechanical work to Haiku-powered subagents. You get the quality where it matters and the cost savings where it doesn’t.
MCP: Claude Talks to Your Actual Systems
Model Context Protocol is the open standard for connecting Claude to external services - your database, GitHub, Linear, Slack, internal APIs. An MCP server exposes tools, resources and prompts in a standard format, and any MCP-capable client can use them.
The reason MCP won as a standard is that you write the integration once. The same server works in Claude Code, Claude Desktop, the API via the MCP connector and other vendors’ tools that adopted the protocol.
Where this beats copy-pasting context into chat: freshness and write access. Claude querying your Postgres schema through MCP is always current. Claude with a Linear MCP server can create the ticket, not just draft the text for one.
If you’re wondering where RAG fits in all this: classic retrieval pipelines still make sense for large document corpora, but a lot of what teams built RAG for in 2024 is better served today by a 1M token context window plus MCP tools that fetch live data on demand. Don’t build a vector database to answer questions a grep tool call can answer.
Docs: modelcontextprotocol.io
The Agent SDK: Claude Code as a Library
The Claude Agent SDK is Claude Code’s harness packaged as a Python and TypeScript library. You get the agent loop, the built-in tools (file read/write, bash, grep, web search), context management, hooks, subagents and permissions - and you call it with a prompt and options object.
This is the fastest path from “Claude Code works great interactively” to “we run this in CI” or “our product has an agent in it.” You’re not rebuilding the loop, the tool execution or the permission model. You’re configuring a harness that already handles them.
One clarification worth making because the naming trips people up: the Agent SDK is not the same as the tool runner in the regular API SDK. The tool runner automates the request-execute-loop cycle for tools you define yourself. The Agent SDK ships the entire Claude Code toolset and harness. If you want a coding agent on your own infrastructure, you want the Agent SDK.
Where to Actually Learn This
Anthropic ships free official training at Anthropic Academy - self-paced courses with hands-on exercises and certificates, no payment or subscription. The developer track covers the API from first call to production, MCP server development in Python and agent architecture. If you’re onboarding a team onto Claude, start there instead of scattered blog posts.
For reference material, the two documentation homes are platform.claude.com/docs for the API and code.claude.com/docs for Claude Code, skills, hooks and the Agent SDK.
How the Pieces Fit Together
The way to think about the whole stack:
| You want | Reach for |
|---|---|
| A reusable procedure or checklist | Skill |
| A rule that must always hold | Hook |
| Isolated or parallel work | Subagent |
| Access to external systems | MCP server |
| An agent in your own product or pipeline | Agent SDK |
| Persistent project facts | CLAUDE.md |
Most teams we talk to are underusing hooks and skills specifically. They’ve written elaborate CLAUDE.md files full of rules the model sometimes follows, when a 10-line hook would enforce the rule every time and a skill would keep the context cost near zero.
Start there: take the three instructions you repeat most often, turn the procedural ones into skills and the non-negotiable ones into hooks. That’s an afternoon of work and it compounds on every session after.