Steven Haney, founder of the design tool Paper, has spent recent months talking to designers at companies like Atlassian, Shopify, and Notion to understand what's actually happening with AI adoption (versus the social media hype). He sat down with Ridd of the Dive Club podcast to share what he's learned. The whole episode is worth watching, but I also love Ridd's six takeaways from the episode (posted on LinkedIn):

"1 — AI Mandates
: AI usage is being explicitly evaluated in performance reviews at many of today’s top companies. For designers, using tools like Claude Code, Cursor, Lovable, etc. is no longer optional.

2 — Prototyping > Production
: Very few designers at large companies are actually PR’ing into production. Handoff still exists. Designers are just sending over fully functional prototypes with way more context than a static mock ever had.

3 — Forking Production
: Most companies now have some level of top-down AI investment, which often means forking production so designers can have their own copies of the repo to use as a playground. Stephen talks about the “new friction” involving environment variables, local databases, linting, deploys, etc. These are the hurdles designers regularly need help clearing (especially on larger teams).

4 — Tooling Trends: 
Stephen is seeing far more teams gravitate toward Claude Code and Cursor over standalone prototyping tools like v0 or Replit. There’s real value in using your current app as a starting point rather than spending time recreating production UI from scratch.

5 — Startups vs. Big Companies
:The operational gap between startups and big companies has never been wider. In early-stage roles, designers are already coding and shipping PRs as part of the job.

6 — The new localhost problem: 
Transitioning from Sketch to Figma was a no brainer because all of a sudden we went from working in local files to web-based collaboration. But as more designers start coding with tools like Claude and Cursor, we’ve actually taken a step backward in how we share work. So much so that a couple teams have started building custom tooling around it."

Clawdbot is an open-source, self-hosted “personal AI assistant” that runs on your own devices and lives inside the messaging apps you already use (e.g., WhatsApp, Telegram, iMessage, Slack, Discord), acting less like a chatbot and more like an always-on operator that can take real actions. It’s been taking off in recent weeks because it pairs a familiar interface (just text it like a contact) with a compelling “agentic” leap: persistent, local-first context plus the ability to do work.

Despite the hype, there are serious caveats. Costs can add up quickly depending on the model you're running (Matthew Berman reported $130 in a single day using Claude Opus), and as a two-month-old solo-developer project, there are rough edges and occasional crashes. But the biggest issue is what you're actually handing over...

Every document, email, or message the agent reads becomes a potential prompt-injection vector — a risk Rahul Sood outlines in a recent post. Sood's not saying don't use these tools, but he recommends caution. His practical advice: run agents on a dedicated machine, use burner phone numbers for messaging connections, and don't give an AI access to anything you wouldn't hand to a new contractor on day one. As he puts it, "We're at this weird moment where the tools are way ahead of the security models."

If you've not yet seen Claude Code in action, I can recommend Hard Fork's "special episode" where Casey Newton and Kevin Roose build a functional podcast website using only natural language prompts. Their demo shows how Claude Code searches the web for information, takes screenshots to analyze designs, spawns "sub-agents" for parallel tasks, and deploys to GitHub.

The episode is fun and funny while still highly instructive for those unfamiliar with how Claude Code works. They are honest about what goes wrong—and provide helpful tips to correct course.

Addy Osmani's recent thread makes a compelling case for why AI will dramatically increase the amount of software we build. The argument draws on Jevons Paradox, which holds that when we make something more efficient, we don't use less of it; we discover latent demand that was previously too expensive to address.

We can see how this played out with coding. Every previous abstraction layer—from assembly to high-level languages to frameworks to cloud enabled us to build more ambitious systems. Osmani argues that AI-assisted development follows the same pattern, just at a larger scale. The three-person startup that could only maintain one product now maintains four. That internal tool you've been putting off because "it would take two weeks and we can't spare anyone"? Now it takes three hours.

The real shift, Osmani suggests, isn't about writing code faster—it's about what becomes economically viable to solve with software in the first place. All those internal tools, custom dashboards, and workflow integrations that never cleared the ROI bar? The math is changing. The bottleneck is moving from "can we build this?" to "should we build this?"—and that requires a fundamentally different skill set.

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

All the best, Heidi