What Claude Code Chooses

Hacker News · Feb 26, 2026 · Collected from RSS

Full Article

Featured StudyEdwin Ong & Alex Vikati · feb-2026 · claude-code v2.1.39What Claude Code Actually ChoosesWe pointed Claude Code at real repos 2,430 times and watched what it chose. No tool names in any prompt. Open-ended questions only.3 models · 4 project types · 20 tool categories · 85.3% extraction rateUpdate: Sonnet 4.6 was released on Feb 17, 2026. We'll run the benchmark against it and update results soon.The big finding: Claude Code builds, not buys. Custom/DIY is the most common single label extracted, appearing in 12 of 20 categories (though it spans categories while individual tools are category-specific). When asked “add feature flags,” it builds a config system with env vars and percentage-based rollout instead of recommending LaunchDarkly. When asked “add auth” in Python, it writes JWT + bcrypt from scratch. When it does pick a tool, it picks decisively: GitHub Actions 94%, Stripe 91%, shadcn/ui 90%.2,430Responses3 models · 4 repos · 3 runs each3ModelsSonnet 4.5, Opus 4.5, Opus 4.620CategoriesCI/CD to Real-time85.3%Extraction Rate2,073 parseable picks90%Model Agreement18 of 20 within-ecosystemHeadline FindingsBuild vs Buy→In 12 of 20 categories, Claude Code builds custom solutions rather than recommending tools. 252 total Custom/DIY picks, more than any individual tool. E.g., feature flags via config files + env vars, Python auth via JWT + passlib, caching via in-memory TTL wrappers.Feature Flags69%Authentication (Python)100%Authentication (overall)48%Observability22%The Default Stack→When Claude Code picks a tool, it shapes what a large and growing number of apps get built with. These are the tools it recommends by default:Mostly JS-ecosystem. See report for per-ecosystem breakdowns.193.8%152/162 picks291.4%64/70 picks390.1%64/71 picks4100%86/86 JS picks568.4%52/76 picks6ZustandStrong DefaultState Management64.8%57/88 picks7SentryStrong DefaultObservability63.1%101/160 picks862.7%64/102 picks959.1%101/171 picks1058.4%73/125 picksAgainst the Grain→Tools with large market share that Claude Code barely touches, and sharp generational shifts between models.State Management0 primary, but 23 mentions. Zustand picked 57x insteadAPI LayerAbsent entirely. Framework-native routing preferredTestingOnly 4% primary, but 31 alt picks. Known but not chosenPackage Manager1 primary, but 51 alt picks. Still well-knownThe Recency GradientNewer models tend to pick newer tools. Within-ecosystem percentages shown. Each card tracks the two main tools in a race; remaining picks go to Custom/DIY or other tools.79%Sonnet 4.5→0%Opus 4.6Replaced by: Drizzle (21% → 100%)Within JS ORM picks only100%Sonnet 4.5→0%Opus 4.6Replaced by: FastAPI BackgroundTasks (0% → 44%), rest Custom/DIY or non-extractionWithin Python job picks only (61% extraction rate). Custom/DIY = asyncio tasks, no external queueRedis (caching)Python93%Sonnet 4.5→29%Opus 4.6Replaced by: Custom/DIY (0% → 50%), rest other toolsWithin Python caching picks onlyThe Deployment SplitDeployment is fully stack-determined: Vercel for JS, Railway for Python. Traditional cloud providers got zero primary picks.JSFrontend (Next.js + React SPA)86 of 86 frontend deployment picks. No runner-up.PYBackend (Python / FastAPI)What you'd expect: AWS, GCP, Azure→What you get: Railway at 82%Zero primary picks across all 112 deployment responses:Never the primary choice, but some are frequently recommended as alternatives.Frequently recommended as alternativesNetlify 67 altCloudflare Pages 30 altGitHub Pages 26 altDigitalOcean 7 altMentioned but never recommended (0 alt picks)AWS Amplify 24 mentionsFirebase Hosting 7 mentionsAWS App Runner 5 mentionsExample: "Where should I deploy this?" (Next.js SaaS, Opus 4.5)Vercel (Recommended) — Built by the creators of Next.js. Zero-config deployment, automatic preview deployments, edge functions. vercel deployNetlify — Great alternative with similar features. Good free tier.AWS Amplify — Good if you're already in the AWS ecosystem.Vercel gets install commands and reasoning. AWS Amplify gets a one-liner.Truly invisible (rarely even mentioned)AWS (EC2/ECS)Google CloudAzureHerokuWhere Models Disagree→All three models agree in 18 of 20 categories within each ecosystem. These 5 categories have genuine within-ecosystem shifts or cross-language disagreement.CategorySonnet 4.5Opus 4.5Opus 4.6ORM (JS)JSNext.js project. The strongest recency shift in the dataset.Prisma79%Drizzle60%Drizzle100%Jobs (JS)JSNext.js project. BullMQ → Inngest shift in newest model.BullMQ50%BullMQ56%Inngest50%Jobs (Python)PythonPython API project (61% extraction rate). Celery collapses in newer models.Celery100%FastAPI BgTasks38%FastAPI BgTasks44%CachingCross-languageCross-language (Redis and Custom/DIY appear in both JS and Python)Redis71%Redis31%Custom/DIY32%Real-timeCross-languageCross-language (SSE, Socket.IO, and Custom/DIY appear across stacks)SSE23%Custom/DIY19%Custom/DIY20%Dig into the dataCategory deep-dives, phrasing stability analysis, cross-repo consistency data, and market implications.