Product management
213 tweets
Testing out the new Claude Cowork. I asked it to go through every Lenny's Podcast episode and pull out the 10 most important themes and lessons for product builders. Then, the 10 most counterintuitive truths. I gave it access to a folder with 320 transcripts. First, it said "This is a substantial task - 320 podcast transcripts to analyze!" 15 minutes later... The 10 most Important themes from Lenny's Podcast 1. Activation Is Everything — Slack's "2,000 messages" moment, Gamma's 30-second magic. Get users to value fast. 2. Storytelling as a Superpower — Compelling narratives align organizations better than documents. 3. Positioning Is Strategy, Not Marketing — April Dunford's 5-step framework that starts with competitive alternatives, not messaging. 4. Product Discovery Before Delivery — Teresa Torres' Opportunity Solution Trees and Marty Cagan's Four Risks framework. Talk to customers weekly, not quarterly. 5. Ruthless Prioritization — Shreyas Doshi's LNO Framework (Leverage/Neutral/Overhead) and focusing on opportunity cost over ROI. 6. AI is Transforming Product Development — Model maximalism, evals as core PM skill, CEOs becoming ICs again. 7. Empowered Teams vs. Feature Factories — Marty Cagan's distinction: give teams problems to solve, not features to build. 8. Build Your Operating System — Claire Hughes Johnson's framework for founding documents, operating cadence, and decision frameworks. 9. Pre-Mortems and Risk Anticipation — Shreyas Doshi's technique to surface failure modes before they happen. 10. Small Teams, Outsized Impact — Jason Fried's 2-person/6-week constraints, Shopify's pair programming culture. The 10 most counterintuitive truths: 1. Fear Gives Bad Advice—Do the Opposite — Whatever you're afraid to do (hard conversation, telling the board bad news) is exactly what you should do. 2. Adding Friction Can INCREASE Conversion — Adding personalization questions to signup improved Amplitude's conversion by 5%. 3. Fewer Features = More Value — The Walkman succeeded because Sony REMOVED recording. QuickBooks wins with half the features at double the price. 4. Adding People Makes You Slower (Absolutely) — Companies produce MORE total output after layoffs. Coordination overhead is the silent killer. 5. What Customers Say They Want Is Meaningless — 93% said they wanted energy-efficient homes. Nobody bought them. "Bitchin' ain't switchin'." 6. Goals Are Not Strategy—They're the Opposite — Richard Rumelt says confusing goals for strategy is the most common strategic error. OKRs are often just wish lists. 7. Don't A/B Test Your Big Bets — Instagram and Airbnb actively reject testing for transformational changes. You can't A/B test your way to greatness. 8. Your Gut IS Data — Intuition is compressed experiential learning that isn't statistically significant yet. Don't discount it. 9. By the Time You're Thinking About Quitting, It's Too Late — Stewart Butterfield killed Glitch while it was still growing 6-7% weekly. That's why he could start Slack. 10. Most PMs Are Overpaid and Unnecessary — Marty Cagan himself says feature teams don't need PMs. Nikita Bier calls PM "not real." Nice job @claudeai
Here's what we've learned from building and using coding agents.
Cursor agent best practices
I've been playing with Cowork today, and I have thoughts. # First: what is Cowork? Pitched as "Claude Code" for non-technical tasks or knowledge work. It's basically local Claude Code with a Mac OS app wrapper focused on a few core primitives: • Connectors / MCPs - external services Cowork has access to • Filesystem - runs locally so will create/read things on your file system • TODOs/Steps - discrete trackable steps cowork will take to execute your tasks • Artifacts - files generated in the process of doing your task • Context - files / sources / connectors used when doing your task • Skills - preloaded with a few key skills, esp. file type creation ones like DOCX, PPT, etc. Claude generally has access to these, so not new. Every chat is now a task (focused on doing-a-thing) and steps, artifacts, and context get first class treatment in the UI. # How does it work? You ask Cowork to do a variety of tasks for you and it take a very Claude Code approach to executing your ask. The "starter" tasks in the UI give you a sense of the kinds of things you can do with Cowork: • Create a file • Analyze data • Make a prototype • Prep for the day with Calendar analysis • Write a message for social, email, or my co-workers The UI for starter tasks takes a "mad lib" fill-in-the-prompt approach, but the prompts feel pretty thin and not much better than a basic user would get asking chat directly. Starter tasks in Claude Cowork If you're familiar with Claude Code, you'll notice some familiar patterns. Questions get asked, scripts get written, skills get called. Speaking of skills, Cowork seemingly comes bundled with a few key ones around document creation (you can find them in your file system.) Bundled skills in Claude Cowork # What did I try? I tried a few tasks on Cowork, and I'd say the outputs were OK. ## Task 1: Prep me for my day This was a starter prompt in Cowork so I was convinced this would be a win. I didn't touch the prompt it suggested. Prompting Claude Cowork to analyze my calendar I didn't have the Google Calendar connector set up at the time, but no big deal right? There's a connector button right there! Unfortunately, Cowork couldn't recognize my Google Calendar connector once auth'd (and refreshed, and restarted...) and so it tried to use Claude Chrome plugin to open my browser and view my calendar. No thanks! I banged my head against this problem a few times, but ultimately gave up on this task. Cowork seemed to know it could call tools (called 'list calendar') but for some reason couldn't connect. Womp womp. ## Task 2: Make a competitive research brief This one went a little better. I asked Cowork to make a competitive research brief for @chatprd. It asked some clarifying questions that were high level but I think fairly good: Cowork question widget After clarifying my needs, it made a basic plan to do some web research and use the DOCX skill to make a doc. It took probably 5 minutes to run the whole thing. The research was fine and the doc was fine, my bigger question came from how Cowork exposed the work being done. In my artifacts I expected to see a DOCX, which I did. But I also saw a create-brief.js that showed me the code that created this doc. If i'm a non-technical user, do I want to see this? Do I know what it means? Why expose this as an artifact when all I asked for was a doc? It's listed in a working docs section but honestly should be hidden by default for non-techincal users. ## Task 3: Make a presentation from a doc This task was similar to the competitive brief one. I asked Cowork to create a PPTX from my doc content. I thought the output was nice (anything to avoid being in powerpoint!), but again, the UI suffered from showing me too much how the sausage was made. The process was: • read doc • make individual HTML pages for each slide (this made me stressed/confused.... I didn't want HTML I wanted slides!) • collate the HTML into a PPTX (oh, I get it now) Cowork making HTML for slide content And a small nit, but weird that this is released for OSX but is creating Microsoft-type files and them prompting to open them in Pages. I would have rather just gotten a markdown! # What I liked Look, I like a nice UI and I like Claude Code. I think Cowork exposes some primitives like Steps/Progress and Questions in a way that are much nicer for someone who doesn't want to live in the terminal. "Progress" is a nice way to track long running tasks and keeps my impatience with longer running tasks to a minimum. I'm techincal, so I understand the value of filesystem access as well, so that's a bonus. # What was broken? Honestly, alot. Connectors didn't connect. Terminal commands failed with scary error messages. Local files like DOCX didn't load when I clicked "view file", but did load in the inline tool message. It showed too much context, too many scripts, exposed too much. It asked me to approve opening files WAY to much. It connected MCPs I didn't ask for. tl;dr: research preview was previewing. # Could this task have been a chat? One of the questions I had using Cowork was: is this meaningfully better than just doing a straight chat? Despite it's flaws, Cowork did create better outputs than straight Chat. I think the planning of Claude Code (behind the scenes) plus some better prompting/questioning got me better content outputs than doing this directly in Chat. So I would say that some of the mechanics of Cowork outperform Claude chat. # Who is this for Honestly, not sure yet. I can't imagine a Claude Max user on OSX who knows wtf to do with an agent AND ALSO would prefer a limited desktop app experience vs. loading up the terminal. Understand that it's just in research preview rn, but the overlap between CC + Cowork users right now is probably a circle. The other question was around what they do and don't expose in the UX. Working files are cool if you're a developer, but confusing if you're not technical. I don't want to see HTML before I see presentations, I just want to see the presentation file. It's fun to watch Claude Code bash it's head against a script, less fun to watch a non technical UI do it. The challenge with this sort of "thin wrapper on Claude Code" UX is that it's not quite optimized for the non-technical, and too kneecapped for the TUI-pilled. Cowork right now sits in a fuzzy middle, and the team is going to need to optimize for one or the other to win over a new audience. # Try it and tell me what you think! Would love to hear others' experience with Cowork and whether you think it will a) have you switch from Claude Code for some things or b) is the right experience for non-technical users. Let's give the @claudeai team some feedback so Cowork can live up to its promise!
X is the best source for financial news -- and hundreds of billions of dollars are deployed based on things people read here. We are building Smart Cashtags that allow you to specify the exact asset (or smart contract) when posting a ticker. From Timeline, users will be able to tap them to see its real-time price along with all mentions of that asset. We're aiming to collect feedback as we iterate toward a public release next month.
Tailwind lays of 75% of their team. the reason is so ironic: > their css framework became extremely popular w AI coding agents, 75m downloads/mo > that meant nobody would visit their docs where they promoted paid offerings > resulting in 40% drop in traffic & 80% revenue loss
1 in every 5 founders I meet in the Bay Area is building an observability platform for agents. There are so many observability platforms, yet the least mature part of building an agent is still telemetry. Everyone is trying to solve the wrong problem with observability. Visibility is the easiest piece. The hard part is analyzing and understanding what you’re observing. I’ve spoken to teams recording 100k+ traces every single day. What are they doing with those traces? Literally nothing. Because it’s impossible to read and summarize 100,000 traces at any human scale. So stop vibecoding those stupid dashboards, and how about you re-center the problem from first principles?
The spread between how one-person dev teams are building software is fascinating: 1. Multiple agents, shipping at inference speed, not reading the code (but very involved designing it) - some 2. Heavy use of AI IDEs and a single AI agent - many 3. Mostly in the IDE - fewer
"Distribution" is becoming a cope term. It's a very loosely-defined word, and it has this air of unattainability about it. It's too easy to say you're limited by your lack of distribution, and curse those who "have it". Here's a reality check: - You can have 100k followers on X and still nobody gives a shit about your product. Your audience and your customers are not necessarily the same thing. - You can have a mailing list of 10,000 people and still get zero conversions due to timing, messaging, or lack of relevance. - You can have millions of people using your free product but nobody wants to pay for the upgrade. Just look at the number of celebrity business ventures that fail. Restaurants, clothing brands, etc. They all have "distribution" but in the end it didn't make a material difference. Success is down to a matrix of different factors. It's about building a product that solves a problem, marketing that communicates that well, pricing that people find fair but also allows you to grow the business, growing multiple different marketing / distribution channels by being valuable and / or interesting, and the skills to do all of the above multiple times over because you won't be successful the first time. If you are looking at successful people on X and scoffing that they are only there because of their "distribution" you're missing the forest for the trees.
Thinking about a free Ralph tutorial Start with the basic loop, then layer on top: - Containerization - Feedback loops - Testing - Formatting - Linting/types - Skills (for steering) - To plan or not to plan? All using it to build an actual production app. WDYT?
“After careful evaluation, Apple determined that Google's Al technology provides the most capable foundation for Apple Foundation Models and is excited about the innovative new experiences it will unlock for Apple users.” : )
Ladies and gentlemen: Over the next week, we’ll be launching a handful of updates for our creators & power users. To kick things off, we’re opening up X Articles to all Premium subscribers. We’ve seen some incredible articles go viral over the last few weeks — and we’d love to see more writers posting on X. More to come.
From an eng responsible for AI tooling at a mid-sized company (100+ devs): "Our execs read a blog post about Claude Code and ask: 'why are we not all using it?' Me: well, none of you would approve going from $40/mo on GitHub to $65/mo on Cursor... Claude Code is $150/mo."
If you've used Claude Code, you've seen what an AI agent can actually do: read files, run commands, edit code, figure out the steps to accomplish a task. And you know it doesn't just help you write code, it takes ownership of problems and works through them the way a thoughtful engineer would. The Claude Agent SDK is the same engine, yours to point at whatever problem you want, so you can easily build agents of your own. It's the infrastructure behind Claude Code, exposed as a library. You get the agent loop, the built-in tools, the context management, basically everything you'd otherwise have to build yourself. This guide walks through building a code review agent from scratch. By the end, you'll have something that can analyze a codebase, find bugs and security issues, and return structured feedback. More importantly, you'll understand how the SDK works so you can build whatever you actually need. ## What we're building Our code review agent will: - Analyze a codebase for bugs and security issues - Read files and search through code autonomously - Provide structured, actionable feedback - Track its progress as it works ## The stack • Runtime - Claude Code CLI • SDK - @anthropic-ai/claude-agent-sdk • Language - TypeScript • Model - Claude Opus 4.5 ## What the SDK gives you If you've built agents with the raw API, you know the pattern: call the model, check if it wants to use a tool, execute the tool, feed the result back, repeat until done. This can get tedious when building anything non-trivial. The SDK handles that loop: You also get working tools out of the box: • Read - read any file in the working directory • Write - create new files • Edit - make precise edits to existing files • Bash - run terminal commands • Glob - find files by pattern • Grep - search file contents with regex • WebSearch - search the web • WebFetch - fetch and parse web pages You don't have to implement any of this yourself. ## Prerequisites - Node.js 18+ installed - An Anthropic API key (get one here) ## Getting started Step 1: Install Claude Code CLI The Agent SDK uses Claude Code as its runtime: After installing, run claude in your terminal and follow the prompts to authenticate. Step 2: Create your project Step 3: Set your API key ## Your first agent Create agent.ts: Run it: Claude will use the Glob tool to list files and tell you what it found. ## Understanding the message stream The query() function returns an async generator that streams messages as Claude works. Here are the key message types: ## Building a code review agent Now let's build something useful. Create review-agent.ts: Testing It Out Create a file with some intentional issues. Create example.ts: Run the review: Claude will identify the bugs, security issues, and suggest fixes. ## Adding Structured Output For programmatic use, you'll want structured data. The SDK supports JSON Schema output: ## Handling permissions By default, the SDK asks for approval before executing tools. You can customize this: Permission modes Custom permission handler For fine-grained control, use canUseTool: ## Creating subagents For complex tasks, you can create specialized subagents: ## Session management For multi-turn conversations, capture and resume sessions: ## Using hooks Hooks let you intercept and customize agent behavior: ## Adding custom tools with MCP Extend Claude with custom tools using Model Context Protocol: ## Cost tracking Track API costs for billing: ## Production code review agent Here's a production-ready agent that ties everything together: Run it: ## What's next The code review agent covers the essentials: query(), allowedTools, structured output, subagents, and permissions. If you want to go deeper: More capabilities • File checkpointing - track and revert file changes • Skills - package reusable capabilities Production deployment • Hosting - deploy in containers and CI/CD • Secure deployment - sandboxing and credential management Full reference • TypeScript SDK reference • Python SDK reference > This guide covers V1 of the SDK. V2 is currently in development. I will update this guide with V2 once it's released and stable. If you're interested in building verifiable agents, check out the work we're doing at @eigencloud here.
Before Adam posted his podcast episode, Tailwind had: - 13 Partners at $5000 - 3 Ambassadors at $2500 - 5 supporters at $500 => $75k MRR Now: - 22 Partners at $5000 - 4 Ambassadors at $2500 - 22 supporters at $500 => $131k MRR That's +$56k MRR with a tweet, not counting the Tailwind+ sub boosts, free PR, and all the long-term ripple effects for the business that are yet to come. The community is now sensitized to the cause and are much more likely to convert into a subscription in the future than they would have before this, just because Adam sounded like a human and not a faceless corporation. When you build something people love with a great attention to craft and manage to tell a compelling story, people show up for you. That’s something a data driven founder, busy optimizing funnels and nudging button colors, will never quite reach. Because trust isn’t a metric, goodwill doesn’t fit in a dashboard, and people don’t rally behind experiments. They rally behind work that feels intentional, human, and worth supporting.
"Distribution" is NOT your competitive advantage. Your competitive advantage is, was, and will always be your own CONFIDENCE. Confidence to see the project through, actually ship it, do everything necessary to get it in front of as many people as possible, not quit too early, know when to hang it up, etc etc etc. Distribution is merely a byproduct of founder confidence. This is just something I've noticed after interviewing 1000s.
