🔵 Uber reveals how AI is transforming prototyping - at a cost

Plus: How to use Claude Routines, Gemini comes to desktop, Square's Managerbot offers a refreshing alternative to Chat boxes

Hi product people 👋,

The token-maxxing trend seems to be a double edged sword for major tech companies; on one hand, it means product teams can ship more than ever, but on the other, it can mean companies spend their AI budgets a little too quickly - as Uber revealed this week.

Plus, find out how to use Claude’s new Routines for product work, why new data makes for some pretty alarming reading for ChatGPT, and a new API that lets you politely say no to requests you’d rather not do at work and elsewhere.

Happy Friday!

Rich

Watch on YouTube | Follow on Substack Notes

---

Key reads and resources for product teams

How to build powerful internal AI productivity tools - how AI-native internal tools like Ramp’s Glass turn scattered workflows, skills, and data into a shared, auto-connected productivity engine that quietly compounds your team’s speed every single day.

Vibe code slop remover - use this prompt to remove bad code from your vibe coded apps and projects and for tech debt identification. (Department of Product)

Awesome design systems - Design system inspirations from popular websites. Drop one into your project and let coding agents build a matching UI. (Design MD)

Agent Skills from Google Director Addy Osmani - Includes a planning and task breakdown, API and interface design and front end engineering design skill - all crafted with Google’s best practices. (GitHub)

Anthropic’s Multi-agent coordination patterns and when to use them - Five multi-agent coordination patterns, their trade-offs, and when to evolve from one to another. (Anthropic)

Org Design in the Age of AI - AI collapses costly “translation layers” between roles like PMs, designers, engineers, QA, middle managers, forcing a shift from relay-race orgs to small, autonomous squads and system-generated roadmaps. (Robonomics)

How to incorporate handmade design styles as a trust signal - In an era of AI-generated-everything, AI-fatigued users want designs that look like they were made by a person. Hand-made designs are a new trust signal. (NN Group)

Google’s PM on how to run evals for Conversational Agents - Prism is an open-source evals tool for Conversational Analytics. Google’s PM Kate Grinevskaja explains how you can use it to run evals for these agents. (Google)

Uber reveals how AI prototyping is changing how they build products - Instead of asking teams to imagine how something might work, PMs or designers could put a concrete, interactive artifact in front of them. Conversations moved quickly from “what is this?” to “is this the right approach?”

But, the mass adoption of AI at Uber has led to the CTO admitting that their 2026 Claude Code budgets have been spent already:

---

New product features and innovation this week

Anthropic released a new version of its desktop app this week which includes a new sidebar along with an integrated terminal, file editing and previews. The new UI certainly has some issues, like the new smaller toggle buttons which I personally find frustrating to use together with some buggy behavior and while some have called the rollout an “affront to software” I wouldn’t go that far.

The more interesting feature from Anthropic for product teams this week, though, is the new feature called Routines.

Routines are Claude Code automations you configure once - giving it a prompt, a repo, and connectors - then set to run on a schedule, via an API call, or in response to an event.

Routines come in three flavors:

1. Scheduled - give Claude Code a prompt and a cadence (hourly, nightly, or weekly) and it runs on that schedule

2. API-triggered - every routine gets its own endpoint and auth token, so you can wire Claude Code into your alerting, deploy hooks, or internal tools

3. Webhook (GitHub) - subscribe a routine to automatically kick off in response to GitHub repository events, with Claude creating a new session for every PR matching your filters

The Routines can also make use of Anthropic’s library of third party connectors which includes tools like Amplitude, Notion, Linear, Slack and others

Examples of the kinds of workflows you can set up with Anthropic Routines

While these are generally targeted to developers, product managers, designers and other non-engineers can also make use of Routines if they want to. Here’s some ideas on potential routines you could set up;

• Track whether the insight from last quarter’s research is still true - Use Notion (research repo) + Mixpanel + Linear. Quarterly, a Routine revisits research findings older than six months and checks current Mixpanel behavioural data against the assumptions those findings supported.

• Design debt surfacing - on a nightly schedule, Claude scans merged PRs for component or style changes, cross-references them against your design system documentation, and opens a ticket whenever implementation has drifted from spec

• Get ahead of churn before it registers in your metrics Use Pendo + Gainsight + Intercom + Slack. A routine monitors weekly for accounts showing a combination of declining feature breadth, increasing support contact frequency, and reduced session depth

• Know when your pricing page is confusing people before they leave - Use FullStory + Intercom + Slack. A routine monitors customer support conversations weekly for pricing and packaging confusion - questions about what’s included, which tier they need, whether they can upgrade mid-cycle.

Gemini adds previews in AI Studio, flexes Antigravity with new Mac release

Google Gemini has added design previews in AI studio. This new feature means that when you’re vibe coding a new app in AI Studio, rather than just landing on a specific, final solution for you, the app will show you a bunch of different themes / directions for you to choose from. So, for example, if you’re building a new app from scratch, it might show you options like “professional polish”, “elegant dark” and “immersive UI”.

Here’s a demo of how this new feature works presented by one of Google’s AI PMs to explain more:

For product teams, this could be helpful in scenarios where you want to brainstorm some potential ideas before landing on a particular design direction or for situations where you want to show a bunch of different potential options to users during user testing.

As well as this, Google also introduced some useful new features across its other products. Chrome is getting Skills, which allow you to save prompts and re-use them later. Plus Gemini has joined the battle against Claude Code with the launch of a new dedicated desktop app.

Google’s CEO Sundar Pichai said that the new Gemini desktop app was built entirely using Antigravity in just a few days with “more features on the way”. With the app, you can use the Option + Space shortcut to pull up a floating chat bubble, where you can ask Gemini questions and share your window. You can also share your browser and use that as context and overall, it’s a neat way to get quick access to the Gemini models on your machine but for now, it feels like more of a flex of Antigravity’s capabilities more than anything else and lacks the ability to edit files locally. Let’s see what new features they add over the coming weeks.

Square’s proactive manager agent moves away from the default chatbox UI; but could it cannibalize its third party ecosystem?

Square is embedding an agentic layer directly into its merchant dashboards that can both surface analytics insights and execute account actions.

It’s called Manager Bot and Block’s head of product says “The big shift from Square AI to Managerbot is really from reactive to proactive…What that means is the primary interface is not a question box. You assign tasks to Managerbot, and that could be based on data, an insight, or a signal from your business.”

If Manager Bot handles campaigns, reporting, and bulk item management natively, what’s the value proposition for the third-party apps sitting on Square’s platform? This could cannibalize its own developer ecosystem, which is a real strategic tension they’ll need to manage - and other companies could be susceptible to the same types of disruptive risks.

---

Tools you can use

Clicky - an AI buddy that lives on your Mac. press a shortcut, ask a question out loud, and it walks you through whatever you’re working on

Say no as a service - an API that lets you politely say no to requests.

ProductBridge - collect customer feedback across multiple platforms.

---

📈 Product data and trends to stay informed

Uber’s CTO Neppalli Naga says ~11% of real, live updates to Uber’s code in its backend systems are being written by AI agents built primarily with Claude Code, up from just a fraction of a percent three months ago.

He also said that despite the rapid of adoption of AI models, he’s not thinking about the slowing hiring of software engineers at this point, saying:

“The vision for me as a CTO is to transform from software engineering to [AI] agent software engineering.”

The number of job postings mentioning agentic AI jumped 2,643% in 2025 to reach almost 19% of all AI related job postings. Python remains the most sought after programming language in AI jobs. The full annual AI Index report from Stanford University has over 400 pages worth of analysis.

Vercel’s CEO told the HumanX conference last week that 30% of the web apps it hosts were developed by AI Agents:

The so-called “token-maxxing” trend is beginning to backfire. Meta has pulled its internal leaderboard ranking engineers by token spend and after a manager at Amazon told engineers to use AI more, one of the engineers faked token usage by making each chat interaction use 10x the number of tokens than were actually required.

On Stripe, 65% of transactions under $50 now happen on mobile. The stats were revealed as part of a new report on digital payments that was published this past week. The report also found that in the US, for purchases between $500 and $2,000, mobile usage rose by 4.1% over two years, digital wallets cut average checkout time to 42 seconds, about 2× faster than manual entry and the UK tops the list of countries most willing to make AI-assisted purchases.

43% of AI-generated code changes require manual debugging in production environments even after passing quality assurance and staging tests.

ChatGPT’s market share on the web has dropped from 77% a year ago to 56% in March.

---

Paid subscribers get the full DoP Substack including: The Knowledge Series for sharpening your tech / AI skills, the AI Prompt and Skills library and DoP Deep dive reports for in-depth analysis to learn lessons from the world’s top tech companies.