Gauri Iyer

Yesterday, I put together a new community research project to study how the different agent platforms implement support for Agent Skills. The questions have direct implications for how we should author and publish skills intended to be portable across platforms. If you're a company publishing agent skills, you should be *very* interested in the results, because the implementation differences could be a source of bugs for your customers. You may not have realized you were signing on for it by putting out a few Markdown files that look like docs, but you have implicitly agreed to support the 25+ agent platform "runtimes" for your skills. That means we need to understand how those platforms diverge. But when I sat down in my kitchen yesterday with my morning coffee, I didn't intend to start a new research project. I was just checking a comment thread where a developer had requested a new feature in the `skill-validator` tool I maintain. How did that feature request turn into a GitHub project where I've created infrastructure to ask community members help me test real skill behavior across 25+ platforms? It all started with the greatest developer nerd-snipe of all time... someone on the internet was wrong. Link to the full writeup in the comments.

26

6 Comments

Boris Cherny’s View on Vibe Coding Limitations (from the Podcast) • Best for throwaway/prototype code: “I do this all the time [vibe coding], but it’s definitely not the thing you want to do all the time.” It’s ideal for non-critical path code, quick experiments, or prototypes that might get discarded. • Lacks maintainability and thoughtfulness: “You want maintainable code sometimes. You want to be very thoughtful about every line sometimes.” AI can produce working code fast, but it often results in messy, hard-to-maintain output (e.g., inconsistent style, over-engineering, or subtle issues). • Requires human oversight for quality: On the Claude Code team, they hold AI-generated code to the exact same bar as human-written code. If it “sucks,” they don’t merge it—they ask the model to improve it. • Preferred approach: Pair programming with the model: For important code, align on a plan first (e.g., using Claude Code’s plan mode), iterate incrementally, review/clean up, or even hand-write critical parts. Boris still manually codes core sections where he has strong opinions (e.g., parameter names). • Models aren’t perfect yet: “The models are still overall not great at coding… this is the worst it’s ever going to be.” Massive improvements are coming, but current limitations mean vibe coding alone isn’t reliable for production. Broader Limitations of Vibe Coding (from Industry Discussion & Karpathy’s Own Reflections) Karpathy popularized the term but has since highlighted drawbacks, even abandoning pure vibe coding for complex projects due to persistent bugs and “slop”: • Produces “slop” or low-quality code — Hallucinations, duplicated logic, bloated/over-engineered solutions, inconsistent architecture → Leads to technical debt and hard-to-debug systems. • Security vulnerabilities — AI can introduce exploits (e.g., insecure dependencies, leaked keys) that humans might miss if not reviewing deeply. • Hard to debug and maintain long-term — Lack of deep understanding means subtle/non-obvious problems pile up; fixing requires trial-and-error. • Not suitable for production/critical software — Fine for weekend projects or demos, but risky for anything with real stakes (e.g., scalability, reliability). • Paradox: Best for experienced coders — Experts can spot/fix issues; novices risk building fragile apps without realizing. In summary, vibe coding is a powerful tool for speed and creativity (especially prototypes), but its limitations make it unsuitable as a full replacement for thoughtful, reviewed coding—especially in professional settings. As Boris puts it, treat AI like a pairing partner: collaborate actively, enforce high standards, and intervene where needed. Models will get better rapidly, narrowing these gaps over the next months/years. https://lnkd.in/e446EFR3

4

What can we measure about accessibility on the web using web archives? Looking at the top 500 domains in the Common Crawl Foundation's February 2026 crawl, Pedro Ortiz Suarez and I found that ~40% of extracted colour pairings fall short of WCAG 2.1 AA thresholds, and only ~20% of sites analysed are fully compliant. TL;DR You can do large-scale accessibility research reproducibly from open crawl archives alone. Also, colours on the web are mostly quite ugly. Blog post: https://lnkd.in/eBnrdGZx arXiv paper: https://lnkd.in/eXXY_qgh Interactive results: https://lnkd.in/eV_xfSaU

30

Jennifer Skeem will be joining THU Digital Health to talk about codesigning videogames that can improve young people's lives. How can the creative industries partner with scientists and users to create engaging experiences that help at-risk youth build self-control and better futures? I'll share our UC Berkeley team's process - and the result, un-TILTED!. Scheduled for March 4–6, this virtual summit gathers experts across healthcare, tech, pharma, and creative sectors to advance digital therapeutics. The focus remains on driving real-world results through a foundation of empathy, evidence, and engagement. Learn more about THU Digital Health here: https://lnkd.in/eGBQhk5N #THUDigitalHealth #THU2026 #UCBerkeley #RRLab #unTILTED #SeriousGames #GameDesign

14

What's the design philisophy between RICV CPUs? Actually it's faily simple. In the 1980s, David Pattison was working at the University of California at Berkley. At the time, he noticed that a lot of people were advocating for more feature-rich instruction sets in CPUs. The ISA (Instruction Set Architecture) is the interface between hardware and software. Their reasoning was that more instructions would make software easier to write. But Pattison saw where this was going: hardware bogged-down in ever-increasing complexity. He advocated for delegating complexity up the stack into software and keeping hardware simple. The idea was that a simpler instruction set would generate more instructions per progam, but you could stay compeitive by executing those instructions faster. His work helped to develop what we now know as RISCV.

26

We asked six UC Berkeley colleagues what’s your favorite digital app or tool? Tara Mason, Univeral Design for Learning Consultant—RTL: I use Speechify, a text-to-speech app. I use it to listen to PDFs, news articles, and more. One thing I really like is the hover feature, which lets me skip to the parts I care about most. Joe Feria-Galicia, Ally and Accessibility Lead—RTL: Color Contrast Analyser, the color contrast checker tool used for accessibility and WCAG recommendations.  1. It’s Free 2. Has a big impact 3. Intuitive interface 4. Works across platforms and with all desktop apps 5. It works well and doesn't crash Andi R., Operations Supervisor, KALX Radio—RTL: I end up using Google Forms in my job more than any other app. I feel like I'm constantly collecting information. Audrey Alforque Thomas, Director of Data and Institutional Research—Undergraduate Education: The first app I open and the last app I use is NYT Games. My mom and brothers share our results with each other. It's not not competitive. Noah Wittman, Program Manager, Strategic Initiatives—RTL Gemini and chatGPT, often to expand and refine my thinking, using prompts such as "What am I missing?" or "What else might I consider?" Timothy Lynch, General Manager, KALX Radio—RTL: Adobe Audition brings worlds of possibility to audio production. Analog audio editing with tape and a razor blade was clumsy and the options were pretty limited.  Please share in the comments: what's your favorite digital app or tool? #RTL #RTLUCBerkeley #UCBerkeley

10

Learning Kendo UI? Here’s Why It’s Worth It! If you’re a web developer wanting to level up your skills, take a look at Kendo UI. It offers: 1) A huge set of ready-made UI components (grids, charts, calendars, dropdowns, and more). 2) Easy integration with frameworks like Angular, React, Vue, and jQuery. 3) Fast development of polished, professional UIs. 4) Great documentation and learning resources. I highly recommend starting with the Kendo UI Grid ,it’s powerful for displaying and manipulating data, with features like sorting, filtering, and export to Excel built-in. Learning Kendo UI can boost your career, especially if you’re working on enterprise or data-driven applications. Below are some great links where you can learn Kendo UI. I’ve also used these resources myself:- https://lnkd.in/gUeeEsWN https://lnkd.in/g47SdPyH https://lnkd.in/gV6bVyEu https://lnkd.in/guSPPyRy https://runner.telerik.io/ https://lnkd.in/gD8eCBq2 https://lnkd.in/ghtarxTe https://runner.telerik.io/ #KendoUI #WebDevelopment #Learning #Frontend #EnterpriseApps

12

Sutter Hill Ventures has been hosting these amazing Speaker Series. Today Henry Modisett shared about his time Designing Perplexity. Here is my summary of the talk! We're in a fast-paced cycle of innovation, where companies are only slightly ahead or behind one another. Technology is evolving faster than we can thoughtfully integrate it, creating a "macro novelty effect". When designing, it's crucial we communicate effectively to help others grasp what's happening. The nature of software engineering has shifted—now it's about designing based on direction, not perfection. Early uncertainty is normal; something doesn’t need to be flawless to launch. Simplify complexity for users. What matters most is consistent progress toward a better product—users remember improvement, not old flaws. Ship quickly, iterate fast. This is the mindset. This is the loop. Keep building.

36

1 Comment

Last quarter at Google, my teammate and I built a Gemini CLI web-accessibility extension that allows Gemini to directly access a running web application, perform an accessibility audit, and then attempt to iteratively fix the violations it finds. Most accessibility tools stop at reporting. We wanted to close the loop. Our Goal: Give Gemini the ability to see your web app—and fix accessibility issues autonomously. What this enables: • Gemini uses tools to inspect your local web app and detect accessibility violations in a WCAG-aligned, structured way • Gemini reasons about each issue and applies fixes incrementally • The audit–fix cycle repeats until violations are resolved or reduced This turns accessibility from a static checklist into an AI-assisted fix loop, embedded directly into the developer’s CLI workflow. The extension is open source. Do check it out at https://lnkd.in/gFmG7MWp, and if you find it useful, do ⭐️ the repo If you’re thinking about accessibility at scale, building AI-native developer tools, or curious where agentic workflows make sense, I’d love to hear your perspective. #Accessibility #DeveloperExperience #AIEngineering #GeminiCLI #WebDevelopment #OpenSource #Google

27

6 Comments

Over the past year, I’ve been spending more time building systems that aren’t just used by people. They’re used by AI. That shift has changed how I think about design. We're used to making interfaces intuitive for humans, but AI is a different kind of user. It doesn’t scan a page. It parses structure. It doesn’t get distracted by clutter. It gets tripped up by ambiguity. It doesn’t need polish. It needs clarity. The more I build AI-enabled systems, the more I think we need a clearer set of principles around this. I’ve started thinking of it as AIX - Artificial Intelligence Experience. Not as a buzzword, but as a practical discipline. If a model is reasoning over schemas, calling tools, and operating inside a limited context window, design decisions matter in different ways. In practice, good AIX tends to look like: - Explicit state instead of implicit assumptions - Deterministic, composable tools - Clear boundaries around context - Well-defined schemas - Designing around known failure modes As agent-driven workflows become more common, I think this way of thinking will matter more. For others building AI-native systems, what principles have you found essential when designing for models as users? 

5

Accessibility Tip: Making Breadcrumbs Inclusive Breadcrumbs help users understand where they are within a website’s structure. But without proper accessibility, they can confuse users relying on assistive technologies. Let’s make sure our breadcrumbs are clear, navigable, and inclusive. How to Make Breadcrumbs Accessible �� Use a <nav> element with aria-label="Breadcrumb" 🔹Structure links using an ordered list (<ol>) 🔹Use aria-current="page" on the last item to indicate the current page 🔹Avoid using visual separators (like slashes) as text they should be added via CSS WCAG Success Criteria It May Fail Without Accessibility 🔹1.3.1 Info and Relationships (Level A) - If semantic structure is missing, screen readers won’t convey hierarchy 🔹2.4.4 Link Purpose (Level A) - If links aren’t descriptive, users won’t know where they lead 🔹4.1.2 Name, Role, Value (Level A) - If ARIA roles and states are missing, screen readers can’t interpret the breadcrumb trail Bad Example <div class="breadcrumbs">  Home / Products / Electronics / Headphones </div> 🔴 No semantic structure 🔴 No accessible labels 🔴 Slashes are read aloud by screen readers Good Example <nav aria-label="Breadcrumb">  <ol>   <li><a href="/home">Home</a></li>   <li><a href="/products">Products</a></li>   <li><a href="/electronics">Electronics</a></li>   <li><a href="/headphones" aria-current="page">Headphones</a></li>  </ol> </nav> ✅ Semantic HTML ✅ Clear navigation structure ✅ Screen reader-friendly Breadcrumbs are small, but they carry big meaning especially for users navigating with assistive tech. Let’s build them right. Hexaware TechnologiesHexaware Technologies : IT Careers #Accessibility #WCAG #Breadcrumbs #InclusiveDesign #DigitalInclusion #A11y #WebAccessibility #UX #UI #Hexaware #ARIA #ScreenReaderSupport

13

I just finished the next stage of Orchestruct focused on building the scheduling engine that turns a list of tasks into a realistic weekly plan. Instead of just keeping tasks in a list, the system attempts to place them into an actual time grid so you can see what your week would look like before committing to it. The engine generates a proposed schedule and you can review it and either approve or reject it so it can be adjusted to your liking. A few things that came out of this stage: • A greedy scheduling algorithm that scores tasks by urgency and priority and fills available time slots • Plan and PlanItem models that track the lifecycle of a proposed schedule • A pure function scheduler isolated from the API and database so the algorithm can evolve without affecting the rest of the system • 4 API endpoints, tests, and a migration One design choice I spent time thinking about was keeping the scheduling logic completely separate from the rest of the application. This part of the system is the most likely to evolve over time, so isolating it makes it easier to experiment with new approaches later. With this stage complete the system can now generate a full weekly plan from a list of tasks. Next step is expanding the scheduling logic and introducing more constraints and preferences. Curious how other people approach planning their work or personal goals. What would make scheduling and prioritizing things feel more streamlined for you?

1

Today, I’m excited to share something I’ve been working on... Previsly—a tool I built to help writers and filmmakers turn screenplays into AI-generated video storyboards / table reads. Most AI video tools just aren’t built for writers working with actual scripts—especially not for longer content like pilots or features. With Previsly, you can: * Upload an entire screenplay (not just a short prompt) and generate a video storyboard. * Edit the output by swapping in new AI-generated images or choosing from multiple styles: cinematic, anime, comic book, and more. If you’re curious, check it out or reach out—I’d love to hear feedback from fellow storytellers. https://previsly.com/ #screenwriting #AI #storyboarding #filmmaking #previs

11

1 Comment

Jolene Parish just left Thinking Machines Lab after a year. She's heading back to OpenAI. It's the latest in a string of high-profile returns, turning the AI talent war into a boomerang effect. This follows co-founders Barret Zoph and Luke Metz, researchers Sam Schoenholz and Lia Guy, and others who have made the same move back. It's a fascinating pattern emerging in the AI space. On one hand, it shows the incredible gravitational pull of established giants. The resources, scale, and mission can be hard to resist, even after trying something new. On the other, it highlights the intense competition. Startups like TM Lab, backed by $2 billion and led by former OpenAI CTO Mira Murati, are magnets for elite talent. They attract stars like coding Olympian Neal Wu and PyTorch creator Soumith Chintala. Yet, retaining them in this hyper-competitive market is the next great challenge. What does this mean? 🔁 Talent is fluid. Loyalty is to the mission and the cutting edge, not just the company. ⚖️ For startups, attracting stars is one victory. Building a culture and project that keeps them is another. 🧲 For larger firms, being a 'boomerang' destination is a powerful advantage, but it requires actively welcoming returns. The real story isn't just poaching. It's a complex dance of ambition, resources, and where people believe they can have the most impact at any given moment. Have you seen this 'boomerang' trend in your industry? #AI #TalentWars #FutureOfWork #TechCareers 𝗦𝗼𝘂𝗿𝗰𝗲꞉ https://lnkd.in/dPUPaKAA

2

Spec-driven development on steroids: GSD for Claude Code Been using GSD (Get Shit Done) with Claude Code for a while now, and I'm genuinely blown away. It just works! The premise is simple: vibecoding produces inconsistent garbage at scale. GSD fixes that by being the context engineering layer that makes AI coding reliable. What makes it special: → Asks questions until it actually understands what you want → Creates atomic task plans with built-in verification → Each task runs in a fresh 200k context window (no context rot) → Every task = one surgical git commit → Multi-agent orchestration: researchers, planners, executors, verifiers The workflow: describe → research → plan → execute → verify → ship Just structured shipping. 🚀 Install with one command: npx get-shit-done-cc Repo: https://lnkd.in/eMPEJGQv Has anyone else been using this? Curious about your experience.

7

Recently submitted AgentIvy to Gemini 3 Hackathon! Built an AI-powered agent that automatically tests web applications for WCAG compliance and performance issues. The Problem: Discovered accessibility requirements late in a DMV CRM project The Solution: AgentIvy clones repos, analyzes components, and suggests fixes using Gemini 3.5 Pro Tech Stack: Java + Spring Boot + Angular + Playwright + Google Cloud Try it out: [https://lnkd.in/gSgz87Qn] #Gemini3 #2026 #Hackathon #WebAccessibility #AI #DevTools #AIAgent

30

Here's a hard problem: figuring out which donors are ready to give, and how much they could give, before they even open their wallet. That's what we're building with Momentum at Virtuous. The tricky part? Nonprofits operate under a constraint most businesses don't face: spending money to raise money feels like taking dollars away from the mission. Every software purchase, every new tool — it's a trade-off that requires serious justification. So when we build products for nonprofits, they need to actually work. No fluff, no buzzwords that sound cool but don't move the needle. Momentum uses engagement patterns, giving history, and behavioral signals to surface your highest-probability donor opportunities in real-time. Think of it as getting a heads-up about who's ready to give before you start manually combing through your database. Learn more: https://lnkd.in/gE2D8H_S #NonprofitTech #Fundraising #TechForGood #Virtuous

31

1 Comment