AI Craftsmanship - January 2026

TL;DR

This week in 30 seconds:

• Context Engineering: Anthropic defines the new skill - it’s no longer about prompts, but about token configuration
• Devin’s Review: 67% PR merge rate, 4x faster - but only 15% success rate on complex tasks
• Microsoft Acquisition: Copilot has lost market share - coding startup acquisition expected

---

Audio Version

---

🎯 Technique of the Week

Image: AI-generated with Pollinations.ai

Context Engineering Replaces Prompt Engineering

Anthropic has published a new guide that’s changing the developer landscape [1]. The central thesis:

“Building with language models is becoming less about finding the right words for your prompts, and more about answering: what configuration of context is most likely to generate our model’s desired behavior?”

What is Context Engineering?

Gartner defines it as [6]: “Designing and structuring the relevant data, workflows and environment so AI systems can understand intent, make better decisions and deliver contextual, enterprise-aligned outcomes.”

Why this matters:

Prompt Engineering	Context Engineering
Focus on HOW to ask	Focus on WHAT to provide
Single interaction	System design
Static	Dynamic

Anthropic’s Three Core Principles:

1. Finite Attention Budget - Every token consumes attention. With n² pairwise relationships, precision decreases as context grows [1]

2. Just-In-Time Retrieval - Don’t preload everything; store identifiers and fetch dynamically [1]

3. Goldilocks Zone - System prompts specific enough for guidance, flexible enough for heuristics [1]

Gartner’s Recommendation: Companies should appoint a “Context Engineering Lead” [6].

Bottom Line: Anyone working with AI agents in 2026 needs to master Context Engineering - not Prompt Engineering.

---

🔥 Top Stories

Devin’s Performance Review: The Real Numbers

Image: AI-generated with Pollinations.ai

Cognition has published Devin’s 2025 review [2]. The numbers are impressive - but with nuances:

The Improvements:

• 4x faster at problem-solving
• 2x more efficient at resource consumption
• 67% PR merge rate (vs. 34% last year)
• 20x faster at security fixes (1.5 min vs. 30 min)

Real-World Case Studies:

Customer	Result
Major Bank	5-10% developer time saved [2]
Nubank	8-12x faster migration, 20x cost reduction [2]
Litera	40% more test coverage, 93% faster regression cycles [2]
EightSleep	3x more data features shipped [2]

The Other Side: Answer.AI tested Devin for a month - result: only 15% success rate on 20 tasks [10].

The Pattern: Devin is a specialist, not a generalist. Best at repetitive tasks like migrations and refactoring - not creative product development [2].

---

Microsoft Will Buy a Coding Startup

The Information predicts [3]: Microsoft will acquire an AI coding startup in 2026.

The Reason: GitHub Copilot has lost significant market share. Current distribution:

Tool	Market Share
GitHub Copilot	~25%
Cursor	~25%
Claude Code	~25%
Others	~25%

Why Acquisition?

• Microsoft’s Inflection deal ($650M) showed how to navigate antitrust [3]
• OpenAI already bought Windsurf (ex-Codeium) for $3B [3]
• For developers who switched, an acquisition is the only path back [3]

Possible Targets: Cursor, Replit, or a smaller startup with strong technology.

---

⚡ Quick Hits

• “Show Me The Money”: Menlo Ventures calls 2026 the year of ROI proof [9]
• 105,000 Software Jobs: Contrary to fears, positions are rising - AI drives demand [7]
• Stack Overflow Trust Drops: For the first time, positive AI sentiments are declining significantly [5]
• Super Agents: IBM predicts multi-agent dashboards as standard [4]
• Vibe → OVP: Software practice evolves from “Vibe Coding” to “Objective-Validation Protocol” [4]

---

🛠️ Tool of the Week

Context Engineering Checklist

Based on Anthropic’s Guide [1] - the key techniques:

1. System Prompt Design

❌ Too vague: "Be helpful"
❌ Too rigid: "If X then Y, else Z"
✅ Goldilocks: "You are a code reviewer.
   Focus on security and performance.
   Ask for clarification when unclear."

2. Tool Design

• Minimal, focused, clearly named [1]
• Rule of thumb: If a human wouldn’t be sure which tool to use, it’s too complex

3. Memory Management

• Compaction: Summarize conversation history [1]
• Structured Notes: Persistent memory files outside the context window [1]
• Sub-Agents: Specialized agents for focused tasks [1]

The Mantra: “The smallest set of high-signal tokens that maximize the likelihood of some desired outcome.” [1]

---

💥 Fail of the Week

The 19% Slowdown Returns

The METR study from 2025 has been confirmed in new analyses [5]:

The Problem: Experienced developers are 19% slower with AI tools on their own code - even though they believe they’re 20% faster.

2026 Update: Stack Overflow’s survey shows declining positive AI sentiments for the first time. The hype backlash is real [5].

Possible Reasons:

• Context-switching overhead
• Overconfidence leads to less review
• AI tools optimized for demos, not real codebases

The Lesson: “Show me the money” applies to developer productivity too - not just enterprise ROI [9].

---

📊 Number of the Week

105,000

Open software engineering positions worldwide in January 2026 [7].

The Context: Despite AI fears and layoffs at tech giants, demand is rising. The driver: AI, ML and cloud require more engineers, not fewer [7].

Most In-Demand Skills:

• AI/ML Integration
• Cloud-native Development
• Cybersecurity
• Context Engineering (new!) [6]

The Warning: Entry-level positions are shrinking. AI is taking over junior tasks. Mid-level with AI skills is the sweet spot [7].

---

📚 Reading List

1. Anthropic: Effective Context Engineering - The definitive guide (20 min) [1]

2. Devin’s 2025 Performance Review - All metrics and case studies (15 min) [2]

3. Gartner: Context Engineering - Enterprise perspective (10 min) [6]

---

📅 Next Week

• CES 2026 Recap: What the AI hardware announcements mean for developers
• Context Engineering Deep Dive: Practical implementation
• AI Craftsmanship February: Focus on agent orchestration

---

🤖 Behind the AI

Generated in: ~45 minutes Sources scanned: 30+ articles Stories found: 12 → 6 selected Validation: 4 Agents (Fact-Check, Devil’s Advocate, Quality Editor, Legal) Model: Claude Opus 4.5 Images: Pollinations.ai (3 generated)

Full metrics

Phase	Metric	Value
Source Collection	WebSearch Queries	6
Source Collection	WebFetch Calls	8
Selection	Stories presented	12
Selection	Stories selected	6
Draft	Words	~1200
Draft	Sources cited	10
Images	Used	3
Audio	Generated	Yes (DE + EN)

---

This newsletter was created with AI assistance. All facts were verified against original sources.

Subscribe to newsletter · Give feedback