Frontier Models

Claude's code interpreter changes the game for everyday work

Anthropic quietly dropped what might be the most practical AI update of the year: Claude can now create and edit Excel spreadsheets, PowerPoint presentations, Word documents, and PDFs directly in its interface. While this may sound mundane, the execution appears to be remarkable.

In this live demo by Nate Jones, Claude produced an 8-tab financial model with working formulas, properly documented logic, and readable headers—the kind of spreadsheet that would typically take an analyst all day to build. The PowerPoint outputs showed actual design sensibility with balanced spacing, proper type hierarchy, and centered elements. Most tellingly, when compared head-to-head with ChatGPT's Agent mode on the same tasks, Claude delivered usable work products while ChatGPT's Agent mode produced unreadable spreadsheets and poorly formatted slides.

At this writing, this functionality is not yet rolled out to Pro accounts (it's currently available as a preview for Max, Team, and Enterprise plan users).

The video is worth watching to gain a sense of just how much this could impact day-to-day work in organizations.

Claude Code Interpreter deep dive: Real workflows + prompts
Watch time: 40 minutes

Official announcement: Claude can now create and edit files
⚡ Quick Read (2 minutes)

Frontier Models

The changing nature of AI collaboration

Ethan Mollick's latest piece marks a pivotal shift in how we work with AI. We're moving from being active collaborators who guide and correct AI (what Mollick calls "co-intelligence") to becoming audiences for AI wizards that conjure sophisticated outputs with minimal input. And if you're wondering precisely what that means, please see the newsletter item above.

Here's the catch: AI systems working as autonomous agents plan and execute tasks without our intervention, and we can't see or control their process. This creates what Mollick calls the "wizard problem"—the better the output, the harder it becomes to verify accuracy or understand the choices made:

"So what do we do with our wizards? I think we need to develop a new literacy: First, learn when to summon the wizard versus when to work with AI as a co-intelligence or to not use AI at all. AI is far from perfect, and in areas where it still falls short, humans often succeed. But for the increasing number of tasks where AI is useful, co-intelligence, and the back-and-forth it requires, is often superior to a machine alone. Yet, there are, increasingly, times when summoning a wizard is best, and just trusting what it conjures."

I strongly encourage reading his essay as a reflection on where we're heading, in practical terms, with AI.

On working with wizards
☕ Medium Read (9 minutes)

AI Product Development

GitHub's Spec Kit brings structure to AI vibe-coding

From the GitHub blog:

"As coding agents have grown more powerful, a pattern has emerged: you describe your goal, get a block of code back, and often… it looks right, but doesn’t quite work. This “vibe-coding” approach can be great for quick prototypes, but less reliable when building serious, mission-critical applications or working with existing codebases. Sometimes the code doesn’t compile. Sometimes it solves part of the problem but misses the actual intent. The stack or architecture may not be what you’d choose."

GitHub's new open source Spec Kit toolkit offers a solution. Instead of treating coding agents like search engines (throw in a vague prompt, hope for the best), Spec Kit introduces a structured four-phase workflow: Specify → Plan → Tasks → Implement. The magic is giving the AI more explicit instructions at each step.

On the surface, GitHub’s new Spec Kit reads like waterfall development. But the behavior is different. Spec Kit treats the spec as a living, single source of truth that’s regenerated into plans and bite-sized tasks whenever intent changes. Waterfall, by contrast, assumes a mostly fixed requirements document, sequential handoffs, and large testing late in the cycle.

Spec Kit is clearly targeted at developers (for example, it works with GitHub Copilot, Claude Code, and Gemini CLI), but the blog post describing the process describes principles anyone can use to improve their AI output.

Spec-driven development with AI
☕ Medium Read (7 minutes)

Enterprise AI

Building enterprise-ready AI prompts

This video from Anthropic's Applied AI team isn't your typical Claude chatbot session. Instead, it's a live demonstration using Claude's API through their Console interface to solve a complex real-world problem: analyzing Swedish car insurance claim forms and accident sketches to determine fault. Unlike casual chatbot conversations, this video shows how to craft sophisticated prompts for enterprise applications that must work reliably the first time, every time.

The team demonstrates how to build prompts iteratively—starting with a basic request that completely misunderstood the task, then methodically adding context, structured data, step-by-step instructions, and output formatting until Claude could confidently analyze complex insurance documents. Particularly notable is the emphasis on preventing hallucinations and structuring outputs for downstream systems.

If you've ever wondered about how AI can actually fit into organizational workflows, this demo provides a clear, helpful real-world example.

Prompting 101 | Code w/ Claude
Watch Time: 25 minutes

🛠️ Lovable prompt directory

Felix Hass has collected his favorite Lovable prompts into a "master directory," so you can just copy and paste what you need. Prompts include:

• Foundation (auth, users, settings)

• Core UX & UI (dashboards, file uploads, realtime)

• Collaboration & Growth (teams, invites, notifications)

• Monetization (Stripe, PayPal, billing)

• Integrations (Slack, Resend, Maps, Calendly)

• Advanced Systems (feature flags, analytics, cron jobs)

• AI Superpowers (chatbots, semantic search, rec engines)

The directory is built for Lovable, but many of the prompts are generic enough to adapt to other tools. (However, others are clearly Lovable-specific.)

Lovable prompt directory

That’s it for this week.

Thanks for reading, and see you next Wednesday with more curated AI/UX news and insights. 👋

All the best, Heidi