The Cursor Pattern, Beyond Code

What the Commit Data Actually Says

In late February 2026, SemiAnalysis published a tracker that read commit signatures on public GitHub pushes. Roughly 4 percent of all public commits carried Claude Code's signature. By early May that share was up to 4.5 percent, with Claude Code generating about 2.6 million commits a week. Cursor does not publish equivalent data, but the broader signal is the same. Daring Fireball put the platform in context: GitHub now processes roughly 275 million commits a week, a 14x jump year over year.

The cleanest reading is not "developers got lazier" or "AI got smarter." A growing slice of commits share a trait. The AI sits inside the editor. The developer never left the file to ask for help.

The Pattern Is Not About Code

The standard explanation for Cursor's growth is that it has a better model. That is the wrong abstraction. The model is table stakes. Cursor, Claude Code, and GitHub Copilot run on roughly the same frontier menu. The differentiator is workflow.

Compare two developers writing the same React component. Developer A is in Cursor: tab through suggestions, ask a chat panel about a function, accept a diff hunk by hunk. Developer B is in vanilla VS Code with ChatGPT in a tab: describe the problem, copy code into the chat, paste the answer back, fix imports that broke, lose track of which version is current. Developer A finishes in fifteen minutes. Developer B finishes in forty. The model gave the same answer to both. The workflow ate the difference.

A compounding effect runs alongside. Developer A is building muscle memory: shortcuts, diff review patterns, prompt habits all become reflex. Developer B builds none of it because the loop is too slow to repeat. The gap widens with hours.

This is the pattern. The AI sits where the work already lives. The expert keeps final judgment. The interface gets out of the way. Cursor did not invent the idea. Cursor proved it scales.

Where the Pattern Keeps Showing Up

Granola lives in the meeting. The product runs on the laptop, captures audio without sending a bot into the call, and writes structured notes the user can edit. In March, Granola raised $125 million at a $1.5 billion valuation; revenue grew 250 percent in the prior quarter, per TechCrunch. The competing pitch, generic chat where you paste a transcript afterward, is losing share even though transcription is a commodity. Granola won by living where the meeting happened.

Anthropic shipped Claude into Microsoft 365 on May 7. Word, Excel, and PowerPoint went generally available; Outlook entered public beta. The four add-ins share one conversation thread per user, per The New Stack. Change an assumption in Excel and the cover memo in Word inherits it. The bet is that paid professional work happens inside Word and Excel, and that the model belongs there too.

We saw the same logic when we started on LeadLord. The question we kept asking was "where does a wealth advisor running ads actually work?" Not in a marketing dashboard. In the Meta Ads UI, the LinkedIn Campaign Manager, the compliance review doc in Word, the CRM, the calendar, the phone. So LeadLord had to live in those places, not replace them. Compliance gets reviewed inside the draft. Ads ship to the platform where they will run. The CRM receives the lead in the format it expects.

Customer support tells the same story: the AIs winning seats inside Zendesk and Intercom are taking share from standalone chatbots. Marketing automation tells it again, with winners inside HubSpot and Salesforce rather than standalone copywriting apps.

Why Horizontal AI Keeps Losing Skilled Work

Four reasons, stacked.

Muscle memory. Experts have already built workflows in their tool. A lawyer's hands know where the styles menu is. A developer's fingers know the keystrokes for find-in-files. A marketer knows the Meta Ads campaign structure cold. Asking them to leave that tool to talk to a chatbot is a context-switch tax on every prompt.

Context. The file is open. The spreadsheet has the assumptions. The campaign has the audience targeting. The CRM has the prior history. An AI inside the tool already has all of that. An AI in a browser tab needs the user to paste it in, which is slow and a privacy decision most users would rather not make.

Audit trail. Specialized tools log into the system of record. Cursor edits a file, git captures the diff. Claude edits a Word document, the firm's DMS retains it. Copy-paste loses provenance.

Final judgment. The human keeps the decision. The lawyer accepts redlines clause by clause. The developer accepts hunks. The advisor approves the ad copy before it ships. Vertical tools surface the model's work in a form the expert can review at the granularity of their craft. Horizontal chat dumps a wall of text the expert has to transplant and verify by hand.

Each advantage is small. Stacked, they are the moat.

What This Means for AI Product Strategy in 2026

Build for one workflow inside one tool. Not for "everyone, everywhere."

The fragmentation of work is the moat. A wealth advisor's day looks nothing like a litigator's day, which looks nothing like a support manager's day. The horizontal players, ChatGPT and the consumer Claude and Gemini, will keep being the default for generic tasks: drafting an email, summarizing an article, throwaway code in a notebook. That floor is large and it is not going away.

The ceiling, the paid specialized work, is leaking into domain tools. The model layer is commoditizing on price and quality. The integration layer is not. Owning the surface where a profession does its work is more defensible than owning a better model. The surface accrues data, habit, and switching cost in ways the model alone cannot.

The strategic question for anyone shipping AI in 2026 is which surface to own. Pick a profession. Pick a tool. Live there. Build for the keystrokes already in the user's fingers.