Adam Wolff

Spec-First Regenerative Development I think we are about to change how software gets rebuilt. For decades the workflow has been straightforward: idea, spec, code. But working with AI tools like Claude Code is pushing me toward a very different pattern. Instead of starting with a greenfield spec, I start by researching existing libraries or systems and then do this. Step 1: Use Claude to deeply understand the library Point Claude at the repository and ask it to describe everything in detail. What is the purpose of the library? What problems is it solving? How is it structured? What assumptions does it make? What are the invariants and design tradeoffs? Have Claude write all of this into a single architectural document. Call it a Claude.md or whatever you like. The key is that it becomes a complete semantic description of the system. Step 2: Use Claude to improve the design in the document This is the design phase. This is where the innovation happens. Once the description exists, ask Claude to critique it. What would it change if the library were built today? Which abstractions are overly complex? What historical baggage can be removed? What should the architecture look like now? You are now working at the design level, not inside messy code. Step 3: Generate a full test suite from the spec Ask Claude to create tests that define correct behavior. Unit tests, property tests, edge cases, failure modes. Then review those tests carefully. The tests become the contract that the new implementation must satisfy. Step 4: Rebuild the library from scratch Now turn Claude loose to implement the library cleanly using only the improved spec and the tests. Modern structure. Clear interfaces. No legacy constraints. If the implementation passes the tests, it is valid. At that point something subtle has happened. The original codebase is no longer the source of truth. The spec and the tests are the source of truth. The code is simply an artifact generated from them. This feels like a new pattern for modernizing software. Instead of spending years carefully refactoring legacy systems, we can use AI to understand them deeply and then re-derive them cleanly. I suspect many of our existing libraries will eventually be rebuilt this way. Not because AI writes better code line by line, but because it helps us capture what the system is supposed to be and then regenerate it properly. Anyone else landing on this way of developing?

46

17 Comments

With Opus 4.6 release, "everyone can design, and code" this is old news now. With Agent Teams running simultaneously and coordinating without human intervention, we've moved past "everyone can code." Now everyone can... wait, no one needs to do anything. Designers and engineers are already obsolete. But now middle management is no longer needed too. CEOs finally have the bandwidth to make every decision themselves.  Ultimate founder mode. But hold on: If agents can translate customer feedback into strategy AND execute with 20 coding agents, do we even need CEOs? And if there are no employees... there's no buying power. No buying power = no customers = no companies. Only code remains. AI churning software for no-one in particular. Why this makes sense: 1. No-one knows 2. Whaat? -- Just so we're clear, since apparently it's hard to tell these days: Yes, this is satire.

10

4 Comments

I’ve long pushed back against the myth that “UX = making things look pretty.” I was recently revisiting Denny Kruep's "Improving Talk," which reminds us that design is about making things useful. From Don Norman’s Design of Everyday Things to our own project work, the lesson is clear: - Anticipate and compensate for errors. - Design for interruptions. - Leverage human creativity with machine precision. I shared my reflections in this post... 👉 https://lnkd.in/g6WvqF63

2

Anthropic just dropped Claude Design, and I'm genuinely hyped. The killer feature for me? A unified design system that Claude uses to build design prototypes, slides, and motion graphics from your existing codebase and design files, and then automatically applies them to every future project. Consistent colors, typography, and components without thinking about it. My workflow: Claude Design → Claude Code → GitHub → Netlify. I'm using it to spin up slide decks and MVP prototypes fast, iterate cheaply (fewer tokens before I hand off to Claude Code), and ship designs that actually look coherent. It's shaved meaningful time off the early ideation phase. Is it perfect? No. I still prefer the drag-and-drop editing of Figma in certain use cases, and there are some UI/UX bugs they'll need to iron out. But this is a research preview, and the trajectory for value is clear. Curious, how are others using it, and what improvements do you want to see first? https://claude.ai/design #ClaudeDesign #AI #AITools #ProductManagement #DesignSystems #Anthropic #ProductStrategy

9

2 Comments

The pessimism around vibe coding misses the point entirely, at least for Design. The personal stack I've adopted; Cursor, Claude, and Vercel with some Wispr Flow in between, is genuinely magical for "How it Works" designers. It reminds me of the Flash days when I learned ActionScript and, for the first time, had command over both design and code. Without this command, designers depend on two factors largely outside their control: 1️⃣ Finding an engineer who's both willing and excited to design-on-the-fly with you (of which I've been blessed with many), and 2️⃣ Working at a company that actually supports Design-Engineering pairing from the start. Both are rare. The gap that today's tools close empowers build-focused designers in a way that fundamentally challenges the status quo for team structure, org charts and PDLCs.

6

A year ago, I don't think I'd have predicted the pendulum swing of where design is mostly happening. Ironically, a sizeable amount of my design work lately has been translating non-deterministic AI workflows into deterministic service systems. 🐤

16

2 Comments

As a hardware founder, how do you know when you're ready to engage with engineering partners to execute your product vision? I've worked in this space as a product designer, and frequently speak to engineering teams and manufacturing partners - one of their biggest frustrations with early-stage startups is ambiguity: - A brief that keeps moving - A product strategy that isn't crystal clear - A murky definition of who a product is actually for and what they need it to do Without these elements firmly in place, executing the product vision is almost impossible and it sets everybody up to fail. Your design team needs to know who they are building for, and how your product will be used. Your manufacturing partners need to know exactly what you want to build! Before committing to technical execution of your tech hardware product, make sure you have the following in place: 1. You've got real evidence that the problem you're solving exists and that there is demand for a solution 2. You understand exactly who you are designing for and you've got a clearly identified initial target market (which might not necessarily be your ultimate ideal market....) 3. You know what the product needs to do in order to satisfy your first customers 4. You know exactly who your first customers are going to be! My role isn't to replace the design team or the factory. My role is to get you ready for them. I help founders clarify their product strategy and validate their demand first, so that when you do engage with development partners, you can move fast and execute efficiently. Nail down your strategy. Then let the experts execute it. I’m always open to chatting with founders who are trying to navigate this early stage. If you want to sense-check your plan and book a free discovery, drop a comment below or send a DM. Question for my network: Keen to hear from other design partners and manufacturers in this space — what is the #1 thing you wish founders had ready before they walked through your door?

10

2 Comments

I wrote something based on thoughts I've been sharing with my team on AI and design craft. Basically argues that when building gets cheap, judgment gets expensive, and the distance between something that looks like a product and something that actually serves the person using it is where designers have always lived. That distance is getting more consequential, not less. #Design #AI #Craft https://lnkd.in/gPdQxEbD

43

Rick Doten’s latest article explores how agents can subtly deviate from their mission—not by mistake, but by intent—a phenomenon called drift. His next paper will dive into how to detect and prevent this using an Ethical Framework, Sentinel-Council Architecture, and Knowledge Graphs. A must-read for anyone working with Agentic AI.

8

1 Comment

Alright, Friday hot take. Last week people lost their minds over Claude Design, even calling it a "Figma Killer" but the truth is that Anthropic can't build multiplayer software worth a damn. Just try to build a plugin with your teammates. You'll end up burning a million tokens to build the whole collaboration harness to iterate on a skill. Trying to brainstorm an idea? Two people can't interact with the same artifact in real time. Claude is great when for what it is, but I'm still going to reach for Figma anytime I have to hash out an idea with a teammate.

19

Tool design is the most important aspect of using agents. This underscores the need for Model Context Protocol - when you want agents to work for your business, you want custom tools, and MCP is how you make that happen.

7

I find a lot of satisfaction in using vibe coding tools to turn what used to just be fun ideas into real outputs people can use. A couple weekends ago before our whole family was taken down with the flu for 8 straight days, Annie Clark and I took a conversation from our Thursday date night about how to reduce the murkiness around discussing design roles and responsibilities in the modern day and created Is Like, a framework for understanding what kind of designer you should look for based on your specific needs. Have a short conversation with our evaluator and Is Like will output a full readout of skills to look for, how they should be distributed, and what kind of title you should be looking for. We went from concept to publishing in about 72 hours before influenza-a prevented us from sharing, but we’re healed and would love to hear feedback.

21

4 Comments

How do you assess design engineer / technologist performance? I don’t see much convo around it, despite the role gaining popularity. At Instacart, our work spans design system components, process and system improvement / automation, AI explorations, and ad hoc support across both design and engineering. The challenge is that this kind of connective, cross-functional work can dilute visible accomplishments. Impact is spread across multiple streams, and sometimes the value doesn’t show up until later...(and another variable being design org vs eng org). Even the designers who code. Will there be more qualitative metrics? Feels like many design engineers will run into this question as the role matures. I also kinda see this as an opportunity to collectively shape up what we do as design eng/tech.

15

5 Comments

"User interfaces are gonna go away." : Eric Schmidt He said it out loud on the Moonshot podcast by Peter Diamandis, but it’s something we’ve believed for a long time at Thesys. Traditional interfaces are dumb and rigid. They were built for linear, pre-determined user journeys. They were not built for systems that think, learn, or respond dynamically. Over the past two years, we’ve been heads down building the first practical implementation of Generative UI. Today, our customers use it with the simplicity of integrating an API. Also, Generative UI isn’t just about better user experience and engagement. It also drastically improves execution efficiency of teams as they - ✦ Skip manually building the frontend, saving months ✦ Release faster and reduce time to market ✦ Maintain and scale the UI easily with every new feature or flow Interested in learning more about Generative UI? Dropping a few resources in comments.

50

3 Comments

I listened to Jenny Wen's episode on Lenny's Podcast this week. "The design process is dead." She's right. And the conversation is long overdue. Engineers spin up scrappy prototypes in hours. Agentic solutions are replacing entire workflows. The 6-month discovery cycle is gone. Vision is now 3 to 6 months out, not 3 to 5 years. Design's role is evolving faster than most design orgs can keep up with. But here's what the podcast didn't talk about. In the rush to move faster, we're still not asking the most important question: Faster towards what, exactly? I've watched teams optimise brilliantly for speed. Ship faster. Iterate faster. Deploy AI faster. And build products that were technically excellent, commercially successful and quietly harmful. To users. To communities. To the planet. Time pressure and market success push us to optimise for the short-term win. And nobody stops to ask what the long-term cost is. That's why I've been developing a concept I call Digitainable Innovation. Digitainable = AI & Agentic Solutions + Sustainability. The practice of designing AI and agentic solutions that create measurable positive impact for users, for business, and for the planet. Not as a constraint. As a competitive advantage. It's not about slowing down. It's about being intentional about what we're accelerating towards. In an era where AI agents can run autonomously and ship in hours, the question of direction has never mattered more. Because if design's new job is to point people in the right direction? We better make sure the direction is worth going. What do you think: is impact-led design the next frontier, or a luxury we can't afford at this speed? 👉 Jenny Wen Lenny Rachitsky: link to the episode in the comments. --- * The academic foundation for this concept was established by the Bonn Alliance for Sustainability Research (2019 to 2022), who coined "Digitainability" as the cross-fertilisation between digitalisation and sustainable development. I'm building on that research to bring it into the real world of AI, product design, and business, where it belongs. #Digitainable #AIbyDesign #DesignLeadership #ProductDesign #AgenticAI #ResponsibleInnovation

10

9 Comments

This is a great, short read from a trusted resource, Adam Kinney. What stood out to me was “Traditional design systems — Fluent, Material, Carbon — were built for a world where you knew the content at design time. That world is gone. 💨” So true.

6

1 Comment

New feature alert. Built this in one day with a little help from Claude design, and Claude Opus. A few bits: - I did not go with Claude design’s design. I used it for a quick inspiring and to get the idea out and in front of me before doing a high fidelity version. It was great and made a great version of what I described, but couldn’t nail the finer details which, in my opinion, are equally critical to anything else. - AI still doesn’t just “get” it from the beginning. You still have to clarify your intent fully and with enough context to get the best possible result. Even really powerful models cannot read your mind. Therefore writing and communication skills are still very important and will continue to be. I know there’s a lot of fear in the design and development world right now. The fact is we can move really fast these days. But it doesn’t matter how fast you go if the direction is not clear and there is no real vision or intent to guide the ship. So try out these tools, see what sticks, what doesn’t. But at the end of the day, don’t be discouraged. You can move fast and make cool stuff, just make sure what you’re making makes a meaningful difference. For me, this calorie tracking feature is to help me stay on track with my health goals. I made it in a way that feels good to me and aligns with my vision. Maybe some people will use it, maybe not, but what’s great is that I can make it the way I want it in a way that’s useful. At the end of the day - that’s what matters! If you want to check out the app, it’s available on the App Store here: https://lnkd.in/eUzz2a_u

3

Played around with Claude Design today. Research preview, so rough edges expected, but the core is genuinely impressive. Describe what you want, get a working design, refine through conversation. It reads your codebase and design files on onboarding, builds a design system from them, and keeps output on-brand by default. The ceiling for non-designers is going to move. Founders, PMs, and marketers who've always leaned on designers for early-stage concepts now have a faster path to something shareable and testable. Not good news for Figma, Canva, etc. in the pre-AI design stack. Still early. But directionally significant. https://lnkd.in/eQx9DHeC

7

4 Comments