đ” Codex's Record and Replay lets you record a workflow and re-use it
Plus: Claude Design's new capabilities, Square's coding agent shipped an entire backlog in a few days, show off your actual skills with LinkedIn's connected apps
Hi product people đ,
This week, OpenAIâs Codex unveiled some helpful new features, including one called âRecord & Replayâ which uses computer use models to record a task and save it as a re-usable skill. Weâll take a closer look at some of the ways you might want to experiment with it.
Plus, Square reveals how its new coding agent shipped an entire backlog in just a few days and how LinkedInâs new connectors feature can help you demonstrate the work you actually do day to day to potential future employers.
Happy Friday and have a great weekend!
Rich
Watch on YouTube | Follow on Substack Notes
Key reads and resources for product teams
What developers can build on Shopifyâs Catalog API and UCP
As part of the Spring â26 Edition, Shopify designers and developers built five shopping apps in a few days. Each running on Catalog API for product discovery and Universal Commerce Protocol for the full commerce journey, they show whatâs possible for developers, designers, and anyone with an idea. See everything that launched. (Shopify Editions*)
How to use Agent Skills - an in-depth new guide from Googleâs AI teams with practical examples
Agent Skills turn any general-purpose AI agent into a specialist on demand. Written by a specialist AI team at Google including Tanvi Singhal and Gabriela Hernandez Larios, this comprehensive guide explains why a simple folder with a markdown file has become the standard for equipping AI agents with procedural memory - and why 19% of poorly-designed skills actually degrade performance. Learn how to build, evaluate, and deploy skills in production, from structuring your first one to managing libraries of hundreds. (Google)
New on the Department of Product Substack this week:
How Coinbase built an AI Agent that converts Figma designs into production ready code
Learn more about how Coinbase cut feature development time from 16 days to 4 by building an AI Agent that converts Figma designs directly into production-ready code - slashing 80% of wasted handoff time between design and engineering.
Why Semantic Data Matters to Product Teams
Semantic data layers - a translation tool between raw data and business questions - are having a renaissance in 2026. Find out why - and how you can use them at work. (Department of Product)
How to make AI agents follow your design system
AI agents are great at learning from your codebase - which is a problem if your codebase is a mess. Alice Moore discovered this when an AI agent opened a pull request that looked perfect in screenshots but used deprecated components, hardcoded colors instead of design tokens, and failed accessibility checks. Learn how to constrain agent behavior through linting, strict types, and reference implementations so your code review focuses on what actually matters. (Builder IO)
What Atlassian learned building its own DESIGNmd file in production
Atlassianâs design team tested DESIGN.md, and found it works great for quick prototyping but can cost more tokens and produce less accurate results than their custom tools in production. (Atlassian)
Is Meta destroying its engineering organization?
Gergely Orosz investigates how leadershipâs obsession with AI led to forced data-labeling assignments, keystroke tracking, mass departures, and a security breach that exposed high-profile Instagram accounts. (Pragmatic Engineer)
Google Deepmind leader on Agent Harnesses and the future of personal software
Logan Kilpatrick, who runs Google AI Studio and the Gemini API, explains why the agent harness - not the model API - is now the connective tissue across every Google product, why he thinks custom scaffolding has roughly 12 months of relevance before models absorb it, and what 350,000 Android apps built in a single week tells you about where personal software is heading. (Sequoia Capital)
*sponsored by Shopify Editions
New product features and innovation this week
Codex has launched a new feature called Record & Replay which lets users show Codex a workflow once and then reuse it later as a skill.
When you perform a task while Codex watches, it records your actions, identifies the pattern, and converts it into a reusable skill. The next time you invoke that skill with new inputs, Codex executes the entire workflow automatically without requiring detailed instructions.
The system works across multiple tools and interfaces. So, for example, Codex can interact with spreadsheets, web applications, browser tabs, and connected plugins - using whatever combination you demonstrated. It remembers not just the mechanical steps but your preferences and conventions.
For product teams, this new feature could be used for all sorts of workflows. Hereâs some ideas on how you could put this to use:
North star metric âmorning briefâ - record your daily routine: open GA4/Amplitude, select the product area segment, export key charts, grab experiment stats from your experiment platform, and update a Notion âmetrics wallâ. Then you can say âGenerate this weekâs metrics brief for Consumer appâ and have it reârun the exact workflow with a new date range.
Crossâplatform consistency check - demonstrate how you compare the same flow across mobile, web, and desktop: open them sideâbyâside, perform key tasks, and make a structured list of deltas you accept vs. deltas you consider bugs. The replayable skill becomes your âconsistency inspectorâ whenever platform teams diverge.
Weekly âstate of productâ briefing - record how you: open your analytics dashboards, latest user research notes, experiment pipeline, and roadmap board; then compile a short status doc and send it via email/Slack. Turn this into âPrepare weekly product brief for [Team]â so the ritual survives even if a specific PM is out.
Codex has also launched a dedicated iOS plugin that makes it easier to build iOS apps end to end. It lets teams automate the repetitive parts of iOS development (boilerplate, build loops, validation) while keeping the process fast and auditable. You can run your app in a built-in browser without leaving Codex and it comes with a bunch of skills, including a âLiquid Glass expertâ. OpenAIâs co-founder calls it a âmuch better way to build iOS appsâ but weâll let you be the judge of that one.
Square unveils âbuilderbotâ - its latest AI coding orchestration layer; engineers shipped an entire backlog in days
Block has released notes on how it built Builderbot, an AI agent orchestration system designed to manage code changes across massive, multi-service codebases. It functions as a Slack-integrated development assistant that coordinates multiple AI agents to handle tasks ranging from bug fixes to cross-service migrations.
While typical AI coding tools work within single repositories, Builderbot maintains context across Blockâs entire codebase - hundreds of millions of lines of code spanning hundreds of services. Engineers tag â@builderbotâ in Slack with a request, and the system researches the relevant code, creates branches, writes implementations, opens pull requests, and iterates based on CI feedback.
Block says that the new coding bot âpicks up tickets directly from Linear and Jira, creates the branch, writes the code, opens the pull request, watches CI, and iterates based on feedbackâ. On a practical level, the product teams at Square say that builderbot âtook a list of features sellers had been waiting on for monthsâ and âshipped them in daysâ.
Jack Dorsey calls it âthe beginning of the beginningâ of the companyâs intelligence layer:
They also make the point that it was engineers, not product managers, who made the decisions that shaped the product in another example of how AI is eating into the traditional boundaries that used to separate roles in product development.
Claude Designâs new bidirectional syncing
Claude Design shipped some major updates worth knowing about this week. The first is a change to the components it creates to help bridge the gap between design and development. Product teams can now upload their actual component libraries - buttons, typography, color tokens, spacing rules - from GitHub, design files, or raw uploads. Claude then generates designs using only those components and auto-corrects output before showing it to users.
This is part of a new bidirectional syncing with Claude Code, which means that designers can sync their codebaseâs actual design system into Claude Design, ensuring prototypes use real components.
The second is a redesigned editor which comes with some new layout controls. These let you drag, resize and align elements directly on the canvas, without having to rely on the conversational interface to make the changes.
You can also now ship from Claude Design into third parties like Replit.
Other updates worth knowing about
Pinterest has teased the launch of âAsk Pinterest,â an experimental standalone app that lets users discover products through natural language conversation instead of traditional visual search.
Pinterest built it as a separate application to test conversational shopping without disrupting the main app. For product teams, this experimental approach to launching new conversational features as separate standalone apps could be something we start to see more of.
Cursor says Bugbot is now over 3x faster, 22% cheaper, and finds 10% more bugs.
Their product roadmap leaked this week but hereâs one feature for your resume: LinkedIn is launching âconnected appsâ - a feature that automatically generates verified descriptions of how you actually use supported software tools and displays them on your profile.
When you link apps like HubSpot, Adobe Express, or Github Copilot to your LinkedIn account, the platform analyzes your real usage patterns and creates specific statements like âcreates and sends segmented email campaigns in HubSpotâs Marketing Hubâ instead of generic skill tags.
This feels like a neat idea thatâs helpful for proving real world skills to potential employers, but I wonder what the unintended consequences of this might be.
đ Product data and trends
Amazonâs shopping assistant Rufus converts at double the rate of traditional shoppers with the average Rufus user spending 40 minutes before converting vs 15 minutes for non-Rufus users. This suggests that AI shopping assistants not only drive engagement but ultimately conversion, too.
The Rufus data on Assistants comes from the latest Sensor Tower report which includes a bunch of nuggets that might be of interest to product teams, including:
ChatGPTâs market share has dropped below 50% for the first time - driven by Google Geminiâs increasing growth
Men are the predominant users of all leading AI apps except for character based apps like Character AI where women make up almost 70% of users
Shopping and SaaS are the biggest categories for ChatGPT ad spend so far
Metaâs head of AI product for AI for Work Emily Dalton Smith is leaving the company. Their app is still quietly growing, though with downloads up 38% month on month.
49% of US adults now report using AI chatbots, including one in four who uses them on a daily basis according to new Pew Research.
Headcount growth at the most AI-exposed companies is outpacing that at the least AI-exposed companies according to new analysis by PWC:
Retool interviewed over 300 different CIOs, CTOs and CSOs and found that half of respondents discovered company data sitting on free-tier cloud services that nobody had ever approved:
93% of CTOs, CIO and CSOs are alarmed about vibe coded apps in their organization - and theyâre more worried about the ones they donât know about. New data from Retool who are relaunching with security as their differentiator.
Stanford launched a new series of reports designed to monitor the economic impact of AI that examines how AI is impacting both companies and individuals.
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 worldâs top tech companies.