Cursor

Build what models can't do yet, then delete the scaffolding when they can. Cursor repeatedly pulled capabilities from three to six months ahead into the present with extra engineering, then removed that scaffolding once models handled it natively. It moved from autocomplete to agents to the cloud and then to Grok Bot.

Roman Ugarte, employee 15 who led growth for two years, says this made users trust Cursor to bring them the next frontier, and distribution and data advantages followed. The guiding question was "how could we pull that stuff to today."

Lenny's Podcast▶ 15:35

A no-permission culture kept startup pace from 15 to over 1,000 people. Roman's two repeated values are 'deleting the product', removing scaffolding even if it upsets some users, and 'just do the thing', fixing what you see and pulling in the resources you need. Cursor scaled from about 15 people to over 1,000 while keeping that pace.

He argues the competitors Cursor started against lost ground through a cultural inability to move fast, not wrong decisions or lack of resources. Quote: "this is not an ask for permission culture."

Lenny's Podcast▶ 6:24

Being the high-end product produced data to train its own models. Cursor was criticized for lacking a moat, but its position as a high-end coding product let it capture reasoning traces and train its own Composer models.

Anish Acharya's takeaway is to ship first rather than wait for a plan that survives MBA and VC scrutiny: "Moats are most often discovered, not designed." Roman Ugarte makes the same point about Cursor, and Max Schoening cites Cursor, with Intercom, as a company running its own tuned models if the gap to the labs stays small.

Lenny's Podcast▶ 15:35

A new product beat bolting knowledge work onto the coding tool. The team chose not to put knowledge work inside Cursor and built Grok Bot from scratch, which became a breakout product while competitors extended their coding surfaces.

A separate product let them control every pixel and hold one vision, instead of adding tabs per form factor, which Roman calls "a shipping your org chart style thing that I think users are reacting negatively to." Coding tools also intimidate non-technical users and carry a brand association.

Lenny's Podcast▶ 33:18

Being the high-end product produced data to train its own models. Cursor was criticized for lacking a moat, but its position as a high-end coding product let it capture reasoning traces and train its own Composer models.

Anish Acharya's takeaway is to ship first rather than wait for a plan that survives MBA and VC scrutiny: "Moats are most often discovered, not designed." Roman Ugarte makes the same point about Cursor, and Max Schoening cites Cursor, with Intercom, as a company running its own tuned models if the gap to the labs stays small.

Lenny's Podcast▶ 21:30

Cursor grew in a crowded category by out-building, not out-marketing. Acharya points to Cursor and Granola as products that grew despite crowded categories and weak durability stories, arguing "Nobody has a growth problem these days, they have a product problem." Cursor reportedly added about a billion dollars in revenue in a single quarter at two years old.

His advice is to ask whether a growth problem is really a failure of imagination, since AI lets you build ambitious products and charge a lot for them. He sees word of mouth on X, YouTube and Instagram as the best third-party network effect now.

Lenny's Podcast▶ 27:35

Cursor runs strong UGC campaigns despite being a B2B brand. Nick Lawton describes Cursor's team as absolutely crushing it with UGC campaigns on the SideShift platform, which surprised the host given it is B2B.

Lenny's Podcast

Being the high-end product produced data to train its own models. Cursor was criticized for lacking a moat, but its position as a high-end coding product let it capture reasoning traces and train its own Composer models.

Anish Acharya's takeaway is to ship first rather than wait for a plan that survives MBA and VC scrutiny: "Moats are most often discovered, not designed." Roman Ugarte makes the same point about Cursor, and Max Schoening cites Cursor, with Intercom, as a company running its own tuned models if the gap to the labs stays small.

Superwall

A free year of Cursor in Lenny's Product Pass drew fraud rings. Lenny's Product Pass launched with a free year of Cursor, Lovable, Bolt, Replit and v0, and fraud rings, mostly in China, found API exploits that spread through student networks.

He worked with Stripe and Substack to shut them down and his engineer barely slept for a week patching holes. Lesson: don't launch a too-generous free offer without fraud defenses. "That was way too good an offer. People came for it, bad people."

Lenny's Podcast▶ 57:50

Cursor grew in a crowded category by out-building, not out-marketing. Acharya points to Cursor and Granola as products that grew despite crowded categories and weak durability stories, arguing "Nobody has a growth problem these days, they have a product problem." Cursor reportedly added about a billion dollars in revenue in a single quarter at two years old.

His advice is to ask whether a growth problem is really a failure of imagination, since AI lets you build ambitious products and charge a lot for them. He sees word of mouth on X, YouTube and Instagram as the best third-party network effect now.

Lenny's Podcast

Boris Cherny joined Cursor for the product, then left in two weeks. The Claude Code lead briefly left Anthropic for Cursor because he liked the product and team, then returned after two weeks because he missed Anthropic's safety mission.

He still used Cursor in May, when Claude wrote about 30% of his code.

Sub Club

Builders move to Cursor from Bolt or Lovable when they need control. Zevi Arnovitz graduated from chatbots to Bolt to Cursor in light mode to the terminal, leaving Bolt when connecting payments became too much. Bolt, Lovable, Replit, Base44 and v0 use the same models but add layers that make decisions like sign-in and database for you.

Cursor and Claude Code give full tool access and the latest model abilities at the cost of making those decisions yourself: "the main difference between all these tools is basically the harness." Lazar Jovanovic says the same limits and workflows apply in Cursor as in Lovable, though he is stronger with exploratory prototyping tools.

Lenny's Podcast

Builders move to Cursor from Bolt or Lovable when they need control. Zevi Arnovitz graduated from chatbots to Bolt to Cursor in light mode to the terminal, leaving Bolt when connecting payments became too much. Bolt, Lovable, Replit, Base44 and v0 use the same models but add layers that make decisions like sign-in and database for you.

Cursor and Claude Code give full tool access and the latest model abilities at the cost of making those decisions yourself: "the main difference between all these tools is basically the harness." Lazar Jovanovic says the same limits and workflows apply in Cursor as in Lovable, though he is stronger with exploratory prototyping tools.

Lenny's Podcast▶ 1:33:20

Cursor's Composer is the fast executor in a multi-model setup. Arnovitz uses Composer to turn a written plan into a working feature in minutes, while Claude plans as dev lead, Codex handles the hardest bugs and Gemini 3 does UI.

His create-plan command writes a markdown file with a TLDR, critical decisions and tasks with status trackers, so he can hand simple backend work to Composer and frontend to Gemini from the same file. Cursor also runs /review and finishes quickly.

Lenny's Podcast▶ 1:12:35

Composer acts as one reviewer in a cross-model peer review. After Claude reviews its own code, Arnovitz runs the branch through Codex and Composer, then a /peer-review command tells Claude other team leads found these issues and it must fix each or explain why it's wrong, repeating until none remain.

Each model catches different things, giving a non-coder a review process without reading code. Claude sometimes replies that an issue has been raised a third time and is by design.

Lenny's Podcast▶ 13:18