🔵 Figma’s CEO says the Design vs Code debate is a "false dichotomy"
Plus: Claude Code's new Dynamic Workflows, Summer reading list, How Dropbox built its internal coding agent
Hi product people 👋,
This week, Figma’s CEO declared that the design vs. code debate is a false dichotomy - and released a product to prove it. We’ll take a closer look at what it does.
Plus, as the dust settles on last week’s Google’s I/O event, there was one announcement that didn’t get a lot of attention that could transform how the web looks. And Claude Code gets a new Dynamic Workflow feature to tackle complex, multi-stage coding tasks end-to-end.
Happy Friday!
Rich
Watch on YouTube | Follow on Substack Notes
Key reads and resources for product teams
New from the Department of Product Substack this week:
Knowledge Series - How to build a Strategy Thought Partner grounded in Product Frameworks
Adobe’s former CPO warns that outsourcing thinking to AI erodes our strategic abilities - but what if you used AI as a thought partner instead? Learn how to build a product strategy tool grounded in 25+ frameworks (Kano Model, Jobs to be Done, Porter’s Five Forces, and more) that helps you stress-test ideas, evaluate features, and sharpen your thinking without replacing it.
AI Library - Identify opportunities for a Slack AI Agent in your company
Using DuoLingo as inspiration, use this to describe your organisation - team size, tool stack, pain points - and it returns a ranked list of Slack AI agent opportunities, each broken down by what triggers it, which systems it needs to pull from, what it posts back, whether it requires write access, and what the failure risk is. (Department of Product)
Insights - How Uber manages what its AI Agents can see and do
When an AI agent on behalf of an engineer opens a pull request to fix a system bug, who gets credit - and accountability? Uber engineers Matt Mathew, Prasad Borole, and Meng Huang tackle a critical gap in AI identity systems: how to trace actions back through chains of delegated AI agents to the original human user. (Uber Engineering)
Interview - Claude Code creator on why his product manager now codes and the future of software development
Anthropic’s Claude Code creator Boris Cherny believes software engineering as we know it is ending (he hasn’t written a line of code in six months). But unlike some doomsayers, Cherny thinks job destruction will be met with massive job creation. He notes that roles are blending together: his product manager now codes, his designer codes, his manager codes and predicts there will be 100x more of these “builders” in the future. (Platformer)
Design - Why your opening prompt is the most important when using Claude Design
Your opening prompt in Claude Design sets the ceiling for everything that follows. A team at Contra Labs analyzed five designers’ sessions on the same luxury resort brief and found the first prompt mattered more than the next fifteen combined. (Contra Research)
Case study - How Dropbox built its internal coding agent, Nova
Dropbox built Nova, an internal platform that lets coding agents handle not just code writing, but the messy operational work that slows engineers down. In a post by Mike White and Kevin Altschuler, they explain why a shared platform beats one-off AI tools, how they keep agents grounded in real validation systems, and what they’ve learned from scaling agents across their entire development lifecycle. (Dropbox Engineering)
Resource - 50+ Claude Skills for designers and product managers
Tommaso Nervegna has curated 50 Claude Code skills you can install in 90 seconds and use on real work today. Examples include a diagnostic that audits project bottlenecks, a conceptual model builder to align teams on what the product actually is, and a design system audit to catch consistency issues before migration.
Strategy - How to build a self-improving company with AI
This talk from YC partner Tom Blomefield reveals how to extract domain knowledge, build recursive systems that learn from their own failures, and what humans are actually for in this world. Read the full breakdown of what’s coming next. (Y Combinator)
Tech - The 4 new types of AI-driven tech debt - and how to prevent them
Vikram Venkat, principal at Cota Capital, breaks down four new forms of AI debt silently sabotaging enterprise deployments and reveals how leading organizations can prevent catastrophic failure before it’s too late. (Venture Beat)
Summer reading list
New product features and innovation this week
Anthropic has released its new model, Opus 4.8, together with a new feature it calls Dynamic Workflows. This lets Claude Code orchestrate dozens to hundreds of parallel subagents to tackle complex, multi-stage coding tasks end-to-end. Instead of Claude processing work linearly in a single pass, it breaks large problems into subtasks, runs them in parallel across independent agents, and verifies results before presenting them to you.
When you trigger a workflow (either by asking Claude directly with the word “workflow” or enabling “ultracode” mode), Claude plans the work dynamically, distributes tasks across subagents, and uses verification agents to check findings from independent angles. Progress saves continuously, so interrupted jobs resume where they left off. This approach handles work that would normally take weeks or quarters in days.
Here’s some practical use cases for product teams to consider:
Codebase-wide bug hunts - run parallel searches across a service, then have independent agents verify each finding before surfacing real issues in a report
Large migrations - port codebases across frameworks, deprecate APIs, or change languages across thousands of files with coordinated parallel execution
Framework rewrites and tech debt - systematically modernize legacy code by mapping patterns across the old codebase, generating replacements in parallel, and iterating until test suites pass
Managing A/B tests - Claude’s head of product Cat Wu says she recently used this new feature to catalogue 100s of A/B test flags and deprecate stale ones. Instead of waiting for Claude Code to investigate each one sequentially, dynamic workflows allowed her to do them all in parallel.
Because it’s all a script under the hood, workflows are also reusable: you can save them, commit them to your repo, share them with your team and/or turn them into a skill. But, be warned: workflows use significantly more tokens than standard sessions, so budget planning is essential.
Figma’s CEO: Design vs Code is a false dichotomy; new update lets non-engineers make code changes for engineers to review
Figma’s CEO says that the “design vs code” debate is a false dichotomy and that the company is working towards a future that includes both freeform design and prototyping with code and shipping to production - in one connected platform.
The data backs up the idea that more designers than ever are writing code and this week, the company unveiled an update to Figma Make that lets it work using local code bases - much like Claude Code or Cursor. The new update allows non-technical members of product teams like designers and product managers to make visual changes to a product’s live code base and open up pull requests for engineers to review.
Here’s a deeper preview of the new abilities from Figma’s product manager and product marketing manager:
The convergence between coding and design tools is very real so instinctively this makes sense. The problem for Figma, though, is it’s getting squeezed from competitors as the traditional boundaries between product roles continue to blur. This week, Replit joined the fight by shipping new updates to its own design tool, Replit Canvas. Cursor released some lightweight design features in December last year and it wouldn’t be surprising if they, too, added further design capabilities.
Google’s new HTML in canvas, shareable Gemini Workspace assets
Now that the dust has settled on last week’s I/O event, there was one announcement that seemed to get lost among the noise. This is the official announcement of an experimental new API called HTML in Canvas.
Traditionally, web developers faced a choice: use the DOM for semantic richness and accessibility, or use canvas for performance and graphics control. This new API aims to bridge that gap. When you nest HTML elements inside a canvas, the browser can render them into graphics textures while keeping the underlying DOM element interactive and connected to browser features.
For product teams, this could lead to innovative new web interfaces and also make it easier for AI Agents to read the contents of page designs that would otherwise be off limits.
As part of the release, Google’s product manager Thomas Nattestad shared a selection of different demos showing the technology in action.
The demos include a 3D book that uses HTML to render the pages, a dynamic billboard in a 3D environment and new typography designs with refractive overlays. They’re fairly primitive demos right now but we’re likely to see some further, more interesting examples emerge as the API trial continues.
This week, Gemini also got some new updates which make it easier to share chats and collaborate. Projects is coming to Business users of Workspace which will make it easier for groups of people to manage and collaborate on distinct projects and generated media will now also be easier to share.
YouTube will auto-detect AI generated videos and label them more prominently
YouTube’s product teams have confirmed that it will now auto-detect and prominently label AI generated content. On Shorts, the label will appear as an overlay and on longform video it will appear directly below the video player, just above the description.
For users who hate AI generated slop, the updates will be welcome news but the CEO of AI video platform RunwayML isn’t a fan - and advocates for the complete opposite. In his view, we’re headed towards a future where all video is AI-generated and in that scenario, labeling videos as AI generated “would be like trying to determine whether water is wet”. Of course, being the co-founder of a video generation platform, he would say that, but does he have a point?
Tools you can use
Limora - Generate any design asset you actually need in minutes.
Supaboard - transform data into strategy. Run queries like “Give me my acquisition, conversion and activation funnels, and tell me where it’s leaking and I’m losing users.”
Taste MCP - lets your taste follow you across different AI tools.
📈 Product data and trends to stay informed
GPT 5.5 has jumped to the top of the leaderboard of a new software engineering benchmark, DeepSWE:
DeepSWE is a new benchmark that measures how well coding agents handle realistic software engineering tasks by using original prompts written from scratch (not recycled from existing commits), spanning 91 repositories across 5 languages with solutions requiring substantially more code changes (668 lines on average) while keeping prompts short and natural.
Accelerating coding benchmarks continue to have a negative effect on junior software engineering hiring but AWS’s CEO says replacing junior engineers with AI is “the dumbest thing I’ve ever heard”.
Claude Code’s creator Boris Cherny seems to agree. He predicts a 100x increase in people writing code or using AI agents to write code, as productivity gains enable companies to expand rather than contract engineering talent:
Early data on ChatGPT’s new ads product is pretty positive. ChatGPT ads currently generate an average click-through rate of roughly 0.68%, according to Similarweb data, placing them between display ads and traditional Search ads.
DuckDuckGo installs are up 30% after Google’s I/O event last week, with some users expressing worries that Google is “force feeding” them AI features they don’t want.
Claude Mythos Preview has scanned >1,000 open‑source projects It has identified an estimated: 23,019 total vulnerabilities, of which 6,202 are estimated high‑ or critical‑severity. Read the latest status update report from Anthropic.
Anthropic’s growth revenue multiple in the first five months of 2026 dwarfs OpenAI’s, but how long can this realistically last? If a decent chunk of these revenues are a result of inefficient token spend, companies will be incentivised to cut spend or find ways to make it more efficient.
Uber’s CTO famously said he’d spent the entire 2026 budget already last month and this week, their COO said AI spend must show ROI for it to be justified.
It’ll be fascinating to see how revenue growth looks in ~12 months and what companies do to manage spend under greater scrutiny.
Mobile Apps launched before 2020 still generate 69% of all subscription revenue and new subscription app launches went from ~2,000/month in January 2022 to 14,700+ by January 2026, largely driven by AI-assisted development according to the latest State of Subscription Mobile apps report by Revenue Cat.
Other stats from the report worth knowing about for product teams:
AI-powered apps generate 41% more revenue per payer than non-AI apps.
ChatGPT normalised $20/month pricing. AI apps now monetize at roughly 2x pre-AI ARPU. Before ChatGPT, most consumer subscriptions topped out around $60/year.
Health & Fitness leads Day 14 revenue per install at $0.48 median - nearly 6x Gaming’s $0.08.
Business apps lead trial-to-paid conversion at 9.1% median - more than 2x Gaming’s 4.4%.
Visa says 1,000 of its employees are now using Replit for product prototyping and development - and it is investing an undisclosed amount in the company.
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.