AI Dev Essentials #23: Claude adv. file creation, Devins valuation & Gemini Select and Ask

Hey Everyone 👋,

John Lindquist here with the 23rd issue of AI Dev Essentials!

This past week has been full of incremental improvements and refinements across many products. Many of these enhancements eliminate competitors simply by integrating features into providers' default toolsets. For example, Claude's file creation and spreadsheet capabilities wipe out entire categories of AI spreadsheet startups. Gemini's "Select and Ask" feature eliminates many design tools (I've always enjoyed Gemini's style and visuals). It's fascinating to watch providers improve their base subscription tools, systematically displacing entire categories of startup products.

My personal focus this week has been on Claude Code workflows, particularly building local agents that transfer to GitHub agents, and developing patterns for scaling agent teams from individual workflows to enterprise deployments. Looking forward to covering this in my Claude Code workshop this Friday.

🔜 Upocoming egghead.io Lessons

I'm still working on my free "Intro to Claude Code" course and looking forward to posting it soon. It's been a lot of fun to put together. I'm also trying to figure out how to best teach intro to AI CLI tools where I can cover broader concepts that apply to Claude Code CLI, Codex CLI, Gemini CLI, and all the various tools coming out.

Honestly, one of the main reasons I prefer Claude Code to the other tools is its ability to customize settings per session run with regard to hooks and overriding the system prompt, and how much control you have over which MCPs and tools it can use. I strongly advocate for isolated agents with custom configurations per session. While the other tools are catching up, Claude Code is really pioneering those approaches. I'm trying to figure out how to teach these concepts while still accounting for the fact that other tools are catching up.

🚀 Major Announcements

Claude Introduces Advanced File Creation and Editing

Anthropic has launched a comprehensive file creation feature for Claude, enabling users to generate Excel spreadsheets, Word documents, PowerPoint presentations, and PDFs directly through conversations.

Key capabilities:

Private Environment Access: Claude operates in a secure, private computer environment where it can write code and run programs to produce files and analyses
Multi-format Support: Create Excel files (.xlsx), Word documents (.docx), PowerPoint slides (.pptx), and PDFs seamlessly with working formulas and multiple sheets
Conversational Interface: Turn natural language requests into fully formatted documents
Plan Availability: Currently available as a feature preview for Max, Team, and Enterprise plans, with Pro access coming in the following weeks
Settings Integration: Enable through "Upgraded file creation and analysis" in Claude settings under Features

(Anthropic Official, Anthropic Help Center, BGR, The Decoder)

I love what Claude is doing for all these non-developer workflows. I spend a lot of time with my wife, who's the PTA president and manages a lot of school organization activities, classes, and plans that we've used AI for. The vast majority of her work over the past year has always required a bunch of extra steps to take it from an initial idea to what she can actually use with others, which are usually just spreadsheets and slideshows. So this is a huge win personally for me and my family.

Cognition Secures $400M+ Funding at $10.2B Valuation

Cognition, the company behind AI coding agent Devin, has raised $400 million at a $10.2 billion post-money valuation, led by Founders Fund with participation from existing and new investors.

Funding highlights:

Lead Investor: Founders Fund heading the round
Existing Investors: Lux Capital, 8VC, Elad Gil, Definition Capital, and Swish Ventures all participating
New Investors: Bain Capital Ventures, Hanabi Capital, and D1 Capital joining the round
Team Expansion: Two early investors, Christian Lawless (Conversion Capital) and Emily Cohen (Neo), joined full-time
Windsurf Integration: Acquisition more than doubled their ARR and accelerated growth by over 30%
Vision: Building autonomous AI coding agents to transform software engineering workflows

(TechCrunch, Bloomberg, Cognition Blog, CNBC)

Beyond the fact that Cognition controls Devin and acquired Windsurf, both of which I've only spent a little time with and found lacking compared to other tools, it's fascinating to see their valuation being so high when I don't really know many people who are even using these tools. But I guess it's still time to be making big bets, and I'm not discounting their potential. If I had to guess, I'd say one of the providers such as Anthropic, OpenAI, or Google will always have the best developer tools due to their control over the base models and having a complete control over the "stack" from start to finish.

Google's Veo 3 API Gets Major Price Cuts and Feature Updates

Google has significantly reduced Veo 3 pricing while adding support for 1080p videos and vertical video formats optimized for mobile content creation.

API improvements:

Dramatic Price Reductions: Veo 3 drops to $0.40/second (down from $0.75, ~47% decrease), Veo 3 Fast to $0.15/second (down from $0.40, ~62% decrease)
1080p Support: Higher resolution video generation now available for 16:9 aspect ratio content
Vertical Video: Mobile-optimized 9:16 aspect ratios for social media content creation
Production Ready: Veo 3 and Veo 3 Fast declared stable for scaled production use in the Gemini API
Performance Optimizations: Enhanced quality and processing capabilities across all video generation tasks

(Google Developers Blog, WebProNews, The Decoder, Neowin)

My experience with Veo is limited to generating fun and silly videos with my kids. And while I've seen people do impressive things, I'll be curious to see how soon these are turned into anything of substance. I know right now there's a lot of value in creative lifelike videos and posting them to social media to get a bazillion views. So much of social media is just "wait, this must be AI." It's ridiculous that my seven-year-old daughter is already calling out kids videos on YouTube for using AI videos, images, and voices, but that's the world we live in.

🛠️ Developer Tooling Updates

Jules Agent Adds Image Upload Capabilities

Google's Jules coding agent now supports image uploads, allowing developers to share visual bugs, UI mockups, charts, and build failures for more effective debugging.

New image features:

Visual Bug Reports: Upload screenshots of frontend issues for immediate context
UI Mockup Analysis: Share design files and wireframes for implementation guidance
Chart and Data Visualization: Upload graphs and charts for analysis and recreation
Build Failure Screenshots: Share error screens for faster troubleshooting
Format Support: JPEG and PNG image formats are currently supported

(Google Blog - Jules Now Available

Jules from Google is one of those sleeper agents (get it? Sleeper "agent"?) that I think could really take off. Most everyone I know is just waiting for Google to release the next version of Gemini for better tool calling. I personally haven't dug into the Gemini CLI too much simply because I don't want to invest time customizing tooling around Gemini 2.5 when it feels like something better is around the corner. I have no inside knowledge or anything, but the Google hype cycle has been pretty strong lately. And trying to get Gemini 2.5 to consistently work on a task for very long has been a struggle for months. I'd be really curious if Google ever decides to launch a true GitHub competitor.

Gemini Canvas Launches "Select and Ask" Visual Editing

Google's Gemini Canvas introduces a visual editing feature allowing developers to click any element in their web application and describe changes without writing code.

Canvas capabilities:

Visual Selection: Click any UI element to target for modifications
Natural Language Edits: Describe changes in plain English rather than code
Live Preview: See updates instantly in preview mode
No Code Required: Make sophisticated UI changes without programming
Element-Level Precision: Target specific components for granular control

(TestingCatalog - Select and Ask Feature, Google Support - Canvas Help)

This new select and ask feature is probably my favorite new AI feature I've seen in a while. Often you just want to have the AI change a single element on the page but respect the styles on the rest of the page, and being able to interactively click and select rather than trying to describe it through text will turn these into amazing workflows. I've been asking every design AI tool I've talked to implement this feature. It's really caught me off guard that the base tooling from Google came out with it first. Now they just have to make this work 10x faster.

Claude Code Router Enables API Flexibility

The Claude Code Router project allows developers to use any API endpoint within Claude Code, opening access to alternative models and providers while maintaining the familiar Claude Code interface.

Router capabilities:

API Flexibility: Route Claude Code requests through any compatible API endpoint
Model Alternatives: Access models like Qwen 3 Coder 480B through OpenRouter and other providers
Cost Optimization: Take advantage of free or lower-cost model offerings
Interface Consistency: Keep using Claude Code while experimenting with different models
Community Driven: Open source project enabling experimentation and customization

(GitHub - Claude Code Router, Claude Code Router Official Site)

I love seeing more models being able to adhere to Claude Code tooling because in the past, most have been targeting OpenAI's APIs, which have worked the best with IDEs like Cursor and other IDEs. But the more we see models being able to use Claude Code, the more useful Claude Code becomes for everyone. As someone who prefers Claude Code and its host of options over the current CLI offerings (which is aboslutely subject to change), it'll be awesome to see a large variety of models come out that support my favorite CLI.

NotebookLM Expands Multi-Language Support

Google's NotebookLM has launched a comprehensive language expansion, introducing Audio Overviews in over 50 languages with customizable output language selection and multi-language source processing.

Language enhancements:

50+ Language Support: Audio Overviews now available in dozens of languages including Spanish, French, German, Hindi, Japanese, and many more
Source Language Processing: Process documents in multiple languages and generate overviews in your preferred language
Customizable Output: Select your preferred language for Audio Overview generation regardless of source material language
Global Accessibility: Making AI-powered research assistance accessible to non-English speaking users worldwide
Quality Consistency: Maintains the same conversational quality and accuracy across all supported languages

(Google Workspace Updates - NotebookLM Languages, Google Blog - NotebookLM 50 Languages)

NotebookLM continues to be an incredible research and learning tool. If you need to dump a bunch of documentation and URLs to dive deeper into a topic, they're really crushing it with the learning experience. I strongly recommend that if there's any sort of topic you're diving into, start a notebook and just dump every reference and markdown file and everything you have into it.

🤖 AI Ecosystem Updates

Claude Mobile App Gains Real-World Integration

Claude's iOS and Android apps now feature calendar and location services integration, enabling contextual assistance that bridges digital and physical worlds.

Mobile capabilities:

Calendar Integration: Access device calendars to check availability, create events, and schedule meetings
Location Services: Provide nearby recommendations and map-based assistance
Cross-Platform: Works across all Claude plans and integrates with native and third-party apps
App Integration: Compatible with messaging, email, calendar, and location services
Privacy Controls: All integrations require explicit user permission

(Anthropic Help Center)

It's curious to see other providers hook into Google services when Google offers Gemini to control calendars, emails, and such. The anarchist in me quitely hopes that Google pulls the plug on their API for external usage and locks the other AIs out. I keep wondering if we'll see OpenAI and Anthropic offer their own suite of docs/sheets/etc as part of the subscription.

Anthropic Resolves Claude Quality Issues with Transparency

Anthropic identified and resolved multiple bugs affecting Claude model performance, demonstrating improved transparency after community feedback.

Resolution details:

Multiple Bug Resolution: Two significant bugs resolved - one affecting Claude Sonnet 4 from August 5-September 4, 2025, and another affecting Claude Haiku 3.5 and Sonnet 4 from August 26-September 5
Community-Driven Discovery: User reports and detailed feedback helped identify and isolate the quality bugs
Quality Commitment: Explicit statement that they "never intentionally degrade model quality as a result of demand or other factors"
Ongoing Monitoring: Continued surveillance for any additional quality concerns
Timeline Transparency: Clear communication about specific dates and affected models

(Simon Willison's Weblog, The Decoder)

There's been this odd uproar and confusion on social media about Claude's behavior over the past month or so. It's this mixture of reality and what seems like thousands of bots rushing online to criticize anytime something goes wrong, specifically saying they're canceling services and frustrated and giving up. It all smells really funny if you take an eagle-eye view of the whole situation. Some people see no degradation, while other people I trust have noticed minor issues, but nowhere near the response from social media. So it feels like there's both a strange disconnect and this amplification happening through bots trying to drive sentiment around Claude. Anyway, the whole situation smells funny to me, but I definitely appreciate the transparency of Anthropic to say that there were bugs and they've been addressed. The interesting bit is that developers will get frustrated with their tools if things go wrong for a single day because it'll set them back so far that they'll completely abandon it for something else (which is something we would never do in the past because tools were much more stable). People would stick to an IDE or terminal or any developer tools for life and have that be part of their identity. It's fascinating to watch the evolution of developers and their relationship with their to tools.

Kimi K2-0905 Model Achieves Enhanced Performance

Moonshot AI released Kimi K2-Instruct-0905 on September 6, 2025, featuring improved coding capabilities, expanded 256K token context length, and better agent framework integration.

Model improvements:

Enhanced Coding: Significant improvements in frontend development capabilities and tool-calling performance
Extended Context: Context length increased to 256K tokens from previous 128K
Architecture: Mixture-of-experts model with 1 trillion total parameters and 32 billion activated parameters
Agent Integration: Better compatibility with Claude Code, Roo Code, and other agent scaffolds
Performance Benchmarks: Demonstrates superior performance in coding benchmarks and real-world agent tasks
Cost Efficiency: Competitive pricing compared to other high-performing models

(Moonshot AI Official, South China Morning Post)

Kimi is one of those models that seems to be striking a great balance of performance and cost. It's definitely on my list of models to check out again soon, since I haven't used it since their original release. My previous struggles were based around their providers and not so much the model itself.

⚡ Quick Updates

Gemini App Now Supports Audio File Uploads

Universal File Support: Upload audio files to Gemini across Android, iOS, and web platforms, addressing the "#1 request" from users according to Google VP Josh Woodward
Audio Processing: Analyze, transcribe, and work with MP3, M4A, WAV, and other audio formats directly
Usage Limits: Free users get up to 10 minutes of total audio length, while Google AI Pro or Ultra subscribers get up to 3 hours

(9to5Google, Android Police)

NotebookLM Introduces Flashcards and Quizzes

Study Tools: Create customizable flashcards and interactive quizzes from any source material, including lecture notes, research papers, or work reports
Explain Function: Both formats include citation-backed explanations that point back to original source material
Learning Enhancement: Features rolled out in September 2025 with customizable difficulty levels and shareable study sets via simple links

(Google Official Blog, 9to5Google)

OpenAI Codex CLI Gets Enhanced Configuration

Reasoning Control: Configure reasoning effort levels using model_reasoning_effort = "high" in config files or command-line flags for maximum thinking depth with GPT-5 models
Command Integration: New /model command and improved alias support, including shortcuts like codex -m gpt-5 -c model_reasoning_effort="high"
Performance Optimizations: Version 0.23 includes 50% increased usage limits for ChatGPT Plus subscribers and enhanced GPT-5 integration

(OpenAI GitHub Repository, DeepNewz AI)

👆 Cursor Corner

Cursor Team Shares Workflow Approaches

Eric Zakariasson (Cursor Dev Rel) has shared workflow methodologies through the Cursor community, focusing on effective AI coding practices.

Workflow insights:

Community Engagement: Active participation in sharing development patterns and best practices
Practical Experience: Insights from real-world usage at Cursor development team
Methodology Focus: Emphasis on systematic approaches to AI-assisted development
Team Knowledge: Sharing internal learnings with the broader developer community

(Cursor Community Forum, WorkOS MCP Demo Recap)

Cursor Custom Slash Commands Available Through Rules Feature

Cursor IDE has implemented custom slash commands through its "Rules" feature, allowing developers to create personalized shortcuts for AI interactions.

Current features:

Custom Commands: Define your own / commands through Cursor Settings → Rules
AI Action Mapping: Map custom commands to specific AI actions and tools
Workflow Integration: Create shortcuts for common development patterns
Settings Access: Navigate to Cursor Settings → Rules to configure custom commands

(Cursor Documentation - Slash Commands, Cursor Community Forum - Custom Slash Commands)

REST Client Integration Enables Runnable API Documentation

The REST Client extension for VS Code (compatible with Cursor) supports creating API documentation workflows using markdown fenced blocks with .http or .rest code snippets, enabling developers to create "runnable documentation."

Workflow benefits:

Living Documentation: API docs that double as executable test files using HTTP request snippets
Editor Integration: Send HTTP requests directly from markdown documentation files
Fenced Block Support: Embed .http or .rest code snippets in markdown for direct execution
OpenAPI Compatibility: Generate REST client files from existing API specifications
Team Collaboration: Share executable API documentation across development teams

(VS Code REST Client GitHub Repository, REST Client Extension Documentation)

🚨 Last Chance for Friday's Workshop! 🚨

Claude Code Power User Workshop

Ready to master advanced Claude Code workflows? Join this Friday's POWER USER workshop.

Next session: This Friday, September 12, 2025 Time: 5:00 AM - 10:00 AM (PDT) Platform: Zoom 🌍 Europe-Friendly Time! That's 1:00 PM in London, 2:00 PM in Paris & Berlin

Convince your boss: https://egghead.io/workshop/claude-code/boss

Read this far? Share "AI Dev Essentials" with a friend! - https://egghead.io/newsletters/ai-dev-essentials

John Lindquist

https://egghead.io