AI-generated illustration: Developer Tools and Code Visualization — Image generated with Pollinations.ai

AI Craftsmanship 5 min read

AI Craftsmanship - December 2025

Sunday, December 28, 2025

ai craftsmanship claude-code skills mcp agentic-coding

This article was researched and written with AI

TL;DR

Thinking Modes: Claude Code has tiered budgets (“think” → “megathink” → “ultrathink”)
Skills Open Standard: Anthropic opens the Skill specification (December 18, 2025)
Claude Code 2.0.74: LSP integration and terminal extensions
MCP Security: Hundreds of unsecured servers found – but also progress
Agentic Coding: Ecosystem stability matters more than language choice

Technique of the Week: Thinking Modes in Claude Code

Claude Code has tiered thinking budgets. [2] Depending on task complexity, you can give Claude more “thinking time”:

Keyword	Budget	When to Use
`"think"`	Standard	Simple tasks
`"think hard"` / `"megathink"`	Increased	More complex problems
`"ultrathink"`	Maximum	Architecture, debugging

Note: Exact token counts are not officially documented. [2] ClaudeLog provides estimates, but Anthropic hasn’t published precise values.

Important: These keywords work only in Claude Code (CLI), not in the web interface or API. [2]

Limitation: Research shows that more thinking budget doesn’t always yield better results. Self-consistency and iterative approaches can often be more effective than maximum compute time. [3]

Claude Code Updates

Version 2.0.74: The Actual New Features

The December update brings: [1]

LSP Integration – Language Server Protocol directly in terminal: [1]

Go-to-Definition
Find References
Hover Documentation

Terminal Extensions – New /terminal-setup support for Kitty, Alacritty, Zed, and Warp. [1] Syntax highlighting toggle with Ctrl+T.

Already Available Features (not new in 2.0.74)

Checkpoints & Rewind (since v2.0) – Claude automatically saves code snapshots: [1]

Esc + Esc  or  /rewind

Plugin System – Skills, hooks, and commands as distributable bundles: [1]

/plugin install superpowers
/plugin marketplace

Skills Become Open Standard

On December 18, 2025, Anthropic opened the Skill specification – 9 days after founding the Agentic AI Foundation (AAIF) at the Linux Foundation. [7]

What does this mean? [7]

Skills are now an open standard alongside MCP
Other platforms can implement Skills
The Skill syntax is documented and stable

Skills and MCP: Complementary, Not Competing

Aspect	Skills	MCP
Strength	Domain knowledge, workflows	Data integration, tools
Setup	<1 minute (Markdown)	30-60 min (server)
Use Case	Processes, CLI tools	APIs, streaming, databases

Important: Anthropic itself says developers will likely use both. Skills for domain expertise, MCP for external data sources and tools. [7]

Deep Dive: Skills vs. MCP: Why Markdown Files Beat Server Protocols

Agentic Coding Best Practices

Armin Ronacher (Flask author) shared his experiences with agentic coding. [4]

Key Insights

1. Ecosystem Stability Over Language [4]

Ronacher prefers Go – but not because Go is “better,” but because of ecosystem stability. Python’s dynamic features (Pytest fixtures, async patterns) often confuse agents.

Nuance: “Basic Python” (Flask, simple scripts) works fine. The problem is complex frameworks, not the language itself. [4]

2. Everything Can Be a Tool [4]

From shell scripts to log files. What matters:

Speed (agent waits)
Error clarity (agent must understand what went wrong)
Protection against misuse (agents make mistakes)

3. The Classic Flow [4]

Research: Read files, understand
Plan: Design solution (with extended thinking)
Code: Implement
Check: Self-review, tests
Commit: Only when everything is green

MCP: Security Warning and Progress

The Problem

Bitsight Technologies (December 2025): Around 1,000 MCP servers are publicly accessible without authentication. [5]

MCP specification recommends OAuth 2.1 [5]
But authentication is optional [5]
Many DIY deployments skip security [5]

The Solution: Managed MCP

Google launched managed MCP servers for its services: [6]

Google Maps
BigQuery
Compute Engine
Kubernetes Engine

Advantage: Google handles security, hosting, and updates. [6] This shows the way: MCP servers should be operated by experts, not as DIY solutions.

Balance

MCP has real strengths:

Context Preservation across multi-step workflows
Reduced Hallucinations through external data sources
Dynamic Capability Discovery

The security problem primarily affects poorly configured self-hosted servers, not the protocol itself. [5]

Tool of the Week: Claude Code Checkpoints

The checkpoint system (since v2.0): [1]

Problem: Fear of “broken code” during experimental changes.

Solution: Claude automatically saves before each change: [1]

Esc + Esc – Quick rewind
/rewind – Checkpoint selection

You can restore code and conversation separately. [1]

Number of the Week

8x

The difference between standard thinking (“think”) and maximum thinking (“ultrathink”) in Claude Code. [2] But note: More isn’t always better – iterative work can be more effective than one long thinking block. [3]

Reading List

Anthropic: Claude Code Best Practices – Official guide
Armin Ronacher: Agentic Coding – Practical experiences
Skills vs. MCP – When to use which tool?
GitHub: Claude Code CHANGELOG – All updates

Next Week

Deep Dive: Custom Agents in Claude Code
Building a Skill Library: From 0 to 10 Skills
MCP Security: Best Practices for Secure Servers

This newsletter was created with AI assistance and reviewed by 4 specialized validation agents. Hero image generated with Pollinations.ai.