Hey, it’s Andreas.
This week, AI didn’t move in Silicon Valley.
It moved in New Delhi.

India kicked off the AI Impact Summit 2026 and pulled in the people who shape the frontier: Sam Altman, Sundar Pichai, Dario Amodei. In case you missed it, here's a breakdown of what happened.

Quick editorial note: this episode turned long (even after cutting). It’s about Claude Skills, which I consider a very important skill. Pun included at no extra cost. Keeping it in. Going forward, this newsletter stays tighter, and the long-form deep dives go to my X account. Follow there so you don’t miss the full threads.

In today’s issue:

• Google puts AI music creation inside Gemini

• A deep dive into "What is the future of design in a post-AI world?" looks like

• IBM drops 4 new Agentic AI learning paths with 15 hands-on tutorials

• And much more

Let’s get started.

Weekly Field Notes

🧰 Industry Updates

🌀 Google releases Lyria 3 and puts AI music creation inside Gemini → Google’s latest music model, now lets Gemini users create 30-second tracks with lyrics from a text prompt or photo. Here are two examples: a techno chant and a techbro lullaby.

🌀 Google ships Gemini 3.1 Pro across Gemini app, API, Vertex AI, and NotebookLM → Google says 3.1 Pro is its new baseline for complex reasoning.

🌀 Google Pomelli Photoshoot makes studio-quality product images free
→ A new Google Labs tool, lets small businesses upload basic product photos and generate professional studio and lifestyle imagery for free.

🌀 Claude Sonnet 4.6 is out and “better” than Opus 4.5 for most real work → If you run simple agents, switch to Sonnet 4.6 to push more throughput per dollar.

🌀 OpenAI reportedly lines up a $100B funding round at ~$850B valuation
→ Historic round, which is not officially confirmed yet.

🌀 Claude Code to Figma via Figma MCP launched → Figma MCP now allows you to code (design) something using Claude Code and then send it to Figma.

🌀 Apple fast-tracks camera AI wearables → Bloomberg reports Apple is accelerating smart glasses, a camera pendant, and camera AirPods so Siri can use real-time visual context via the iPhone.

🌀 Pentagon may tag Anthropic as a “supply chain risk” over Claude restrictions → Safety policy is colliding with procurement reality, and vendors will be forced to pick a line.

🌀 Accenture ties leadership promotions to weekly AI tool usage → The firm is reportedly tracking senior staff logins and making “regular adoption” a visible input in promotion decisions, forcing AI usage from nice-to-have into a career KPI.

🌀 Ex-Google DeepMind researcher David Silver reportedly raises $1B seed for Ineffable at $4B valuation → If confimed = Europe’s largest seed round to date.

🎓 Learning & Upskilling

📘 DeepLearning.AI course: Gemini CLI for MCP-powered agent workflows → In 73 minutes, you learn to run an open-source CLI agent that uses MCP and extensions to automate dev tasks and coordinate tools like Google Workspace and Canva.

📘 IBM drops 4 new Agentic AI learning paths with 15 hands-on tutorials
→ A structured, practical onramp for building and operating agents, with enough labs to move from concepts to workflows.

📘 Anthropic releases 9 free beginner Claude tutorials → A zero-paywall, non-technical onramp for students and professionals to learn practical Claude usage fast, without needing engineering context.

🌱 Perspectives & Research

🔹 YC’s Lightcone goes inside Claude Code with creator Boris Cherny → An insightful interview on how Claude Code was designed, what actually drives 10x workflows in practice, and where coding agents are heading next.

🔹 ”What is the future of design in a post-AI world?” → Cursor design lead Ryo Lu (ex-Notion) breaks down how Cursor blends design and engineering to empower “designers who build,” and how collaboration changes when AI becomes part of the production workflow.

🔹 Anthropic shows agents are getting more autonomous in real usage
→ Long Claude Code sessions nearly doubled from under 25 to over 45 minutes in three months, auto-approve rises from ~20% to 40%+ with experience, and software engineering still makes up ~50% of agent activity.

🔹 DeepMind CEO Demis Hassabis lays out his “next frontier” AI predictions at India AI Summit 2026 → A broad keynote framing where DeepMind thinks AI is headed next, with emphasis on scientific breakthroughs, healthcare impact, and the governance requirements for responsible deployment at national scale.

🔹 IBM’s agentic AI security guide → Argues that agent threat modeling must be behavioral as well as technical, and recommends monitoring, containment, lifecycle security, and securing the action layer. 

🔹 Princeton University proposes 12 metrics for AI agent reliability - and finds gains are lagging → A new performance profile decomposes reliability into consistency, robustness, predictability, and safety, and tests 14 agentic models on two benchmarks, showing capability gains have translated into only small reliability improvements.

♾️ Thought Loop - What I've been thinking, building, circling this week

If you're a regular reader, you already know I'm a heavy Claude user. And yes, I keep talking about it.

Over the past months I've been going deep on Claude Code, Claude Cowork, and Claude's agentic capabilities. Skills are the next piece of that puzzle, and in my experience, they might be the most underrated one.

I even believe that skills will be bigger than MCP.

And I'll tell you exactly why by the end of this article. But first, let me explain what they actually are - because the confusion out there is real.

What's a Skill? (The Simplest Possible Explanation)

A skill is a note you leave for Claude. Written in plain English. Saved as a file. That's it.

It's not code. It's not a plugin. It's not an API integration. It's just writing clear instructions in plain English, saved in a structured folder.

Skills are instruction packages that give Claude specialized knowledge it does not have built in. Claude knows a lot. But it does not know your company's brand guidelines or how you write your newsletter. It does not know the specific way your team formats financial reports. It does not know the workflow your industry uses for compliance documentation.

Skills fill those gaps. They provide four types of knowledge:

1 ➜ Step-by-step workflows that tell Claude exactly how to complete a process in order.

2 ➜ Domain expertise that gives Claude the rules and standards for your specific field. E.g. what is acceptable in healthcare documentation. Or how legal contracts should be structured.

3 ➜ Tool integrations that teach Claude how to work with specific file formats the right way. Not just creating an Excel file, but creating one with working formulas and proper formatting.

4 ➜ Reusable resources including scripts, templates, and reference docs that Claude can pull from when needed. Your actual templates, not generic ones.

One critical distinction: skills are not just fancy prompts. They are structured packages that persist across conversations and include actual files Claude can work with.

Skills vs Prompts vs MCP vs Subagents: The Confusion Ends Here

People keep mixing these up. Let me clear it once.

Prompts are what you type into the chat. They disappear after the conversation ends. Good for quick, one-off requests.

Skills are packaged procedural knowledge - instructions, code, and assets bundled in a folder. They persist across conversations. They load dynamically. Best for specialized expertise you want Claude to apply consistently.

MCP (Model Context Protocol) gives Claude a live connection to external tools and data sources. It does not teach Claude how to do anything - it gives Claude access to systems. Think of MCP as the data pipe.

Subagents delegate entire tasks to separate AI agents. They carry full agent logic and work across sessions. Best for complex, multi-step tasks that need their own decision-making.

The simplest way to think about it:

1 ➜ Prompts tell Claude what to do right now 

2 ➜ Skills teach Claude how to do something repeatedly 3

➜ MCP connects Claude to where the data lives 

4 ➜ Subagents let Claude hand off work to other agents

They are not competing features. They are layers. A well-built workflow might use MCP to pull data from your CRM, a skill to analyze it using your company's methodology, and a subagent to distribute the results.

Why This Actually Matters (And Why Almost Nobody Gets It Yet)

Three problems plague almost everyone who uses Claude (or other LLMs) regularly.

The consistency problem. Ask Claude the same question on Monday and Tuesday, you might get two different answers. Different structure. Different depth. Different approach. Skills lock in consistent outputs because Claude follows the same instructions every time.

The quality problem. Claude gives decent outputs, but it misses things you know. Industry best practices. Your team's specific standards. The nuances that separate good from great in your field. Skills teach Claude what you have learned over years.

The efficiency problem. You waste time re-explaining context every conversation. Your role. Your preferences. Your constraints. Skills remember so you do not have to repeat yourself.

If you start using Claude Skills, you’ll get streamlined results immediately. You can also share skills with others on your team to make outputs consistent and repeatable. This is especially helpful when you need to follow brand guidelines, apply industry best practices, or keep formats consistent - like writing a newsletter in the same structure every time.

Create a skill once, use it forever. Share it with your team. Everyone gets the same quality output.

A Real Example: Before and After

Let me make this concrete.

Without a skill:

You are a consultant. Every week, you take client meeting notes and turn them into a structured executive summary. You open Claude and type something like:

"Here are my meeting notes. Please turn them into an executive summary. Use this format: a 3-line TL;DR at the top, then key decisions made, then open action items with owners and deadlines, then risks flagged. Keep it under one page. Use professional tone. No bullet points longer than two lines."

Claude does a decent job. But next week you do the same thing. And the week after. And sometimes you forget to mention the format, and the output comes back differently.

With a skill:

You create a skill once called "client-summary." It contains those exact instructions. From that point forward, you just say:

"Here are my meeting notes from the Siemens call. Create the client summary."

Claude recognizes the task. Loads the skill. Follows your format perfectly. Every time. No re-explaining.

Inside a Skill: The Full Anatomy

Every skill lives in a folder. Understanding the structure makes everything else much more easy.

my-skill/
  SKILL.md          <- the only required file
  scripts/           <- code Claude can run
  references/        <- docs Claude reads for context
  assets/            <- files Claude uses in the output

Let me break down each piece.

The SKILL.md File (Required)

This is the brain of the skill. It has two parts.

Part 1: The frontmatter - YAML metadata at the very top, wrapped in triple dashes.

---
name: client-summary
description: >
  Creates executive summaries from client meeting notes.
  Use when user says "client summary," "meeting summary,"
  "summarize the call," or uploads a meeting transcript.
---

The description is the single most important line in your entire skill. It tells Claude when to activate. If your description is vague, the skill will not trigger when you need it.

"Creates Word documents with tracked changes for legal review" beats "Creates documents" every single time.

Part 2: The body - clear, step-by-step instructions in Markdown.

# Client Summary Skill

## Core Rules
1. Always extract action items with owner names and due dates
2. Summarize decisions made, not just topics discussed
3. Flag any unresolved questions that need follow-up
4. Keep summaries under 500 words unless user specifies otherwise

## Output Format
- Executive summary (3-5 sentences)
- Decisions made (bulleted list)
- Action items (table with owner, task, due date)
- Open questions (bulleted list)

## Anti-patterns
- Don't include timestamps unless specifically requested
- Don't summarize small talk or off-topic discussion
- Never assign action items to "the team" - get specific names

That is it. Plain text. No code required.

The Scripts Folder (Optional) holds Python or Bash code Claude can actually run. Use scripts for tasks that must work exactly the same way every time. A script that validates Excel formulas before delivering. A script that checks document formatting against your standards.

The References Folder (Optional) keeps your SKILL.md lean. Move detailed info here: API documentation, database schemas, lengthy style guides, industry regulations. Claude pulls from these files only when relevant.

The Assets Folder (Optional) holds files Claude uses in the output, not for reading. Templates, images, fonts, boilerplate code.

Think of SKILL.md as the manager giving instructions. Scripts, references, and assets are the tools the worker uses to get the job done.

How Skills Work Behind the Scenes

Skills do not load everything into memory at once. That would be wasteful. Instead, they use a system called progressive disclosure.

When you send a message:

1 ➜ Claude scans all available skill descriptions. Just the name and description - roughly 100 tokens per skill. Lightweight. Just enough to know what is available.

2 ➜ If your request matches a skill's description, that skill activates. Claude reads the full SKILL.md instructions.

3 ➜ If those instructions reference additional files - a script, a template, a reference doc - Claude only loads those when it actually needs them.

This means you can build comprehensive skills with detailed instructions, templates, and scripts without worrying about overloading Claude's context window. It only pulls in what the current task requires.

The description field controls everything. If it does not clearly explain when to trigger, the skill sits unused no matter how good the instructions are.

Built-In Skills You Can Use Today

Anthropic already ships five professional skills: DOCX (tracked changes that actually work), XLSX (spreadsheets with validated formulas), PDF (read, merge, split, fill forms), PPTX (presentations that look designed, not generated), and Frontend Design (web interfaces that don't scream "AI made this").

These handle the common stuff, but the real power comes from creating your own.

Build Your First Skill Tonight: Step by Step

Step 1: Pick Your Problem. Ask yourself: What task do I explain to Claude repeatedly? What knowledge do I have that Claude does not?

Step 2: Create the Folder. Make a new folder with your skill name in kebab-case. Something like meeting-notes-processor or brand-content-creator. Only add subfolders if you actually need them.

Step 3: Write the Frontmatter. At the top of your SKILL.md file:

---
name: meeting-notes-processor
description: >
  Transforms meeting transcripts into structured action items
  and summaries. Use when user uploads meeting notes, asks to
  process a transcript, or requests meeting summaries with
  action items.
---

Step 4: Write the Body. Start with what matters most. Critical rules go at the top. Include anti-patterns - telling Claude what NOT to do is often more valuable than telling it what to do. Include examples of input and expected output. Showing beats telling.

Step 5: Add Supporting Files (If Needed). Scripts go in /scripts. Reference docs go in /references. Templates go in /assets.

Step 6: Package It. A .skill file is just a ZIP with a different extension. ZIP the folder and rename from .zip to .skill.

Step 7: Upload and Test. Upload to Claude via Settings > Features > Custom Skills. Test with realistic use cases, not generic prompts. "Test my skill" tells you nothing. "Summarize these meeting notes from the Q3 planning call" tells you everything.

Your first Skill won’t be perfect. That’s fine. Use it, spot the gaps, refine the instructions, re-upload, repeat.

You can use up to 8 Skills in a single request, and you can stack them to get extra leverage.

You can also ask Claude to help you write your Skills. There’s a built-in feature that does an excellent job. That said, I recommend writing your first Skill yourself so you really understand how the pieces fit together.

How to Set It Up: Three Paths

Path 1: Claude.ai or Claude Desktop (Easiest)

This is the path for most people. No code. No terminal.

1 ➜ Open Claude.ai or Claude Desktop

2 ➜ Go to Settings > Features

3 ➜ Make sure "Code Execution and File Creation" is turned on

4 ➜ You will see pre-built skills already active: PDF, Word, Excel, PowerPoint

5 ➜ To add your own: upload your .skill file under Custom Skills

Available on Pro, Max, Team, and Enterprise plans.

Path 2: Claude Code (For Developers)

Drop your skill folder into .claude/skills/ in your project directory. Claude Code detects it automatically with live change detection - you can edit the skill while a session is running and it picks up the changes without restarting.

Path 3: The Claude API (For Builders)

You need three components: the Code Execution tool (sandbox), the Files API (upload/download), and the Skills API (creating/managing skills). The critical detail: skills, code execution, and file handling each require their own beta header. You need all three in every API call. Miss one and it breaks.

The DeepLearning.AI course "Agent Skills with Anthropic" walks through the full API lifecycle. It’s free and the best structured walkthrough I have found.

Mistakes That Will Trip You Up

After spending a lot of time with Claude Skills and seeing what typically goes wrong, here are the things to avoid.

Content mistakes: Explaining things Claude already knows (waste of tokens). Writing vague descriptions that never trigger ("helps with data stuff" will never activate). Cramming too much into SKILL.md - if your file is over 500 lines, split it.

Structure mistakes: Putting files in wrong folders. Forgetting the frontmatter or formatting it incorrectly. Using spaces or uppercase in skill names. Including junk files like .DS_Store.

Trigger mistakes: Putting "when to use this skill" in the body instead of the description field. Claude looks at the description to decide whether to activate. Instructions in the body only get read after activation.

Testing mistakes: Not testing scripts independently before packaging. Testing with generic prompts instead of realistic use cases. And celebrating after one successful test instead of trying edge cases.

One important caveat: community skills on GitHub are hit or miss. I would avoid downloading skills that have not been vetted through careful reading of the markdown files. There are obvious security risks, but there is even greater risk that the skill is just bad and you end up with worse results than if you had just used Claude without the skill at all. Read the SKILL.md before you install anything.

The Bigger Picture: Why Skills Will Be Bigger Than MCP

Here is where I need you to pay close attention. Because this is the part that most commentary is missing.

Anthropic published Agent Skills as an open standard in December 2025. The spec lives at agentskills.io. A skill built for Claude can theoretically work with any AI platform that adopts the standard. You can take a skills folder, point Gemini CLI or Codex at it, and it works - even though those tools have no built-in knowledge of the skills system. I’ve tested this myself with Codex: I pointed it at my existing Skills folder, and the output quality improved immediately - same structure, same rules, less rework.

Here is how I think about it:

MCP standardized how AI connects to your tools. Skills standardize how AI applies expertise inside those tools.

Those are two different layers. Access and judgment. Most enterprise AI deployments are stuck on the first one - connecting systems, building data pipes, setting up integrations. Skills are the second layer. The one that turns a connected AI into a reliable operator.

And here is the part that makes me most bullish on skills: MCP required developers. Skills require domain experts.

You do not need to write code to build a skill. You need to write clear instructions. That is a skill set most professionals already have. There are a lot more domain experts out there than developers.

That is exactly why I think this will be bigger than MCP.

Therefore, the best time to start writing your own Skills - and go deep on the topic - is right now. Here are a few resources to help you go deep into it.

Resources to Go Deeper (ALL FREE)

Further resources (17 free):

OFFICIAL DOCS

• The Unofficial Guide: https://lnkd.in/eNC5QsJp

• Best Practices Guide: https://lnkd.in/emxu8Vsr

• Skills Documentation: https://lnkd.in/eSzfnUNc

• API Reference: https://lnkd.in/erjGW9q5

• MCP Documentation: https://lnkd.in/ejKJuNEX

BLOG POSTS

• Introducing Agent Skills: https://lnkd.in/enrM2tWr

• Engineering Blog: https://lnkd.in/eRn5aYyQ

• Skills Explained: https://lnkd.in/e8zEX2Fe

• How to Create Skills: https://lnkd.in/eDaug-WJ

• Skills for Claude Code: https://lnkd.in/eQpjSyBW

• Frontend Design Skills: https://lnkd.in/efPCkgWb

EXAMPLE SKILLS

• Anthropic's Official Library: https://lnkd.in/er2tG4ZB

• Partner Skills Directory: https://lnkd.in/ejUcTPjT

COMMUNITY LIBRARIES

• Skills.sh: skills.sh

• SkillsMP: skillsmp.com

• Smithery: smithery.ai/skills

• SkillHub: skillhub.club