
A code hosting and collaboration platform for software developers, with free and paid plans.
GitHub's pricing page redesign (more contrast between tiers, more features in higher tiers, the features customers value most emphasized, an outdated tier removed) lifted free and paid signups by over 20%. Its pull request shows the value of a tiny core, and about 10% of top contributors were designers. Lesson: sharpen tier differences, and prototype in code rather than writing specs.
Free and paid signups both rose over 20%
From GitHub pricing page redesign test
A clearer, higher-contrast pricing page lifted free and paid signups by over 20%. GitHub redesigned its pricing page to increase visual contrast between tiers, showing more features in the higher tiers. It also emphasized the features customers value most, used vibrant colors and removed an outdated tier. Free and paid signups both rose by more than 20%.
The changes shipped together, so the gain is not split by change. Contrast made the value gap between plans visible at a glance. Features were chosen to match user priorities rather than listed equally. Dropping the outdated tier left fewer choices to weigh. The lift reached the free tier as well as the paid ones, so a clearer pricing page can help the whole funnel.

GitHub product reviews ran on demos, not memos. Max Schoening says to build a janky demo instead of writing a PRD for the first 10% of a project. That first 10% is now nearly free, and a demo gives people something to react to early, the opposite of waterfall.
He suggests sending off 10 agents to explore 10 paths and seeing which was right. The caveat: the last 10% is still 90% of the work.
GitHub designers prototyped in code to master their material. Around 10% of the top contributors to GitHub's product were designers, even before LLMs. The aim was not production code but understanding the material, such as how an agent loop works.
Schoening would rather have a designer who understands agent loops than one who can only tweak styles. Production contributions tend to follow once people are awakened to the medium.
Mature-product teams should stop being precious about shipping. After a hit first act, teams worry about upsetting users. Schoening tells them users get more upset if you do not innovate than if you accidentally break something, and that the people who built act one can build act two.
He applied this at GitHub and at Notion, and argues cheaper experimentation now lets teams increase shots on goal.
The pull request is a tiny core that is exceptionally good. Schoening names the GitHub pull request among products built on one tiny superpower, alongside iPhone multitouch, Notion blocks, Heroku's git push and Dropbox's menu bar sync icon.
His advice is to find that core and push everything else back, since adding one more thing to finally make a product great never works.
Actions shipped without good package management, and the ecosystem paid for it. Schoening points to GitHub Actions as a case where a missing foundation has cost the ecosystem.
Claude Code accounts for 4% of public GitHub commits. Boris Cherny, head of Claude Code, cites this figure and calls it a floor because private repositories are not counted.
Its pricing page redesign increased visual contrast between tiers, highlighted more features in higher tiers, emphasized the features customers value most, used vibrant colors and removed an outdated tier. Free and paid signups both rose by more than 20%.
A clearer, higher-contrast pricing page lifted free and paid signups by over 20%. GitHub redesigned its pricing page to increase visual contrast between tiers, showing more features in the higher tiers. It also emphasized the features customers value most, used vibrant colors and removed an outdated tier. Free and paid signups both rose by more than 20%.
The changes shipped together, so the gain is not split by change. Contrast made the value gap between plans visible at a glance. Features were chosen to match user priorities rather than listed equally. Dropping the outdated tier left fewer choices to weigh. The lift reached the free tier as well as the paid ones, so a clearer pricing page can help the whole funnel.

GitHub dogfoods every feature internally for months before release, with finance, legal and HR also using it. It started from a developer time problem, then added AI, and funds ideas from customers, open source and pitches. Surveys show code written 55% faster, 85% more confident in quality, 88% less frustrated and code reviews 15% faster. Over 37,000 organizations and 1.5 million developers use it, and it was designed to fade into the background with no friction. Its success is measured by how many secrets were stopped from leaking and how many code issues were fixed before shipping. It came from the same experiment-and-fail approach as Copilot.
GitHub product reviews ran on demos, not memos. Max Schoening says to build a janky demo instead of writing a PRD for the first 10% of a project. That first 10% is now nearly free, and a demo gives people something to react to early, the opposite of waterfall.
He suggests sending off 10 agents to explore 10 paths and seeing which was right. The caveat: the last 10% is still 90% of the work.
GitHub designers prototyped in code to master their material. Around 10% of the top contributors to GitHub's product were designers, even before LLMs. The aim was not production code but understanding the material, such as how an agent loop works.
Schoening would rather have a designer who understands agent loops than one who can only tweak styles. Production contributions tend to follow once people are awakened to the medium.
Mature-product teams should stop being precious about shipping. After a hit first act, teams worry about upsetting users. Schoening tells them users get more upset if you do not innovate than if you accidentally break something, and that the people who built act one can build act two.
He applied this at GitHub and at Notion, and argues cheaper experimentation now lets teams increase shots on goal.
The pull request is a tiny core that is exceptionally good. Schoening names the GitHub pull request among products built on one tiny superpower, alongside iPhone multitouch, Notion blocks, Heroku's git push and Dropbox's menu bar sync icon.
His advice is to find that core and push everything else back, since adding one more thing to finally make a product great never works.
Actions shipped without good package management, and the ecosystem paid for it. Schoening points to GitHub Actions as a case where a missing foundation has cost the ecosystem.
Grew out of the GitHub Next research team and reached general availability about a year and a half after initial product work. Suggestions write roughly 40% of code for Python developers and 27% to 40% across languages, with inline gray text and about 200 ms latency tuned to protect flow. Keeps a separate moonshot team, GitHub Next, apart from its engineering, product and design teams. Ryan splits large team capacity into 5-10% bold bets, about 60% incremental work and 25-30% operations.
Claude Code accounts for 4% of public GitHub commits. Boris Cherny, head of Claude Code, cites this figure and calls it a floor because private repositories are not counted.
Product reviews ran on demos not memos, and about 10% of top contributors were designers before LLMs. Its pull request is a tiny great core, while Actions shipped without good package management, which has cost the ecosystem.
Claude Code reached 4% of public GitHub commits, which Cherny calls a floor because private repositories are not counted.
The Codex name first powered GitHub Copilot years ago. Autocomplete is the model for mixed-initiative AI: cheap when wrong, a speedup when right.
LinkedIn builds its own layer so Copilot, Cursor or Windsurf can work over its code base, and worked with vendors in alpha mode. Its knowledge agents run on Copilot Enterprise, heavily customized.
Cited as the example of early models that only supported autocomplete, before smarter models enabled agentic tools.
Uses an ensemble of models fine-tuned across 30 countries and all their languages, iterating in a loop for next-edit suggestions and code completions. You can already assign a pull request to Copilot.
GitHub began with seat pricing, and Cursor and others have since moved to hybrid pricing. The natural next step is an AI developer or AI QA person that is attributable and autonomous.
It is listed among coding agents that can search the web and so be steered by a malicious page.
GitHub competes as a full system of repo, context, autocomplete, chat and an agent board, which is one of its fastest loops with strong positive feedback. Chennapragada built her new-tab Chrome extension in it in about 10 minutes.
Its first beta was the first truly useful AI product the founders used, and it inspired Cursor. Truell says it stalled because this market does not favor incumbents, and most of the original team has moved on.
Lovable created a new GitHub project per user, and at 15,000 projects a day an on-call engineer cut its access with only an email as notice. Lovable now syncs with GitHub, which Osika says no other tool does this way.
Launched on a small, imperfect model (GPT-2, as Krieger recalled) and was still valuable because it saved typing and users could edit the output. It is the example for human-in-the-loop design.
Its published studies are suggested as the likely analogy for the productivity gains ChatGPT users get on engineering work.
Developer toolsA chip company whose GPUs and CUDA software power gaming, AI and accelerated computing.
Developer toolsAnthropic's agentic coding tool that lets Claude work on code end to end.
Developer toolsAn AI code editor that lets engineers build software with AI while staying in control.
Developer toolsAn AI app builder that lets people create software by chatting with an agent.
An app development company that builds apps for founders and clients.
Developer toolsA fast, well-designed project and issue tracking tool for software teams.
Developer toolsGoogle's backend platform for building apps.
Developer toolsA cloud platform for building, deploying and scaling front-end and AI web applications.