If you're using an AI coding agent, you may not be choosing your own tech stack anymore.
A research team at Amplifying.ai ran Claude Code against 2,430 real repo prompts — open-ended questions like "add auth" or "add feature flags" — and recorded exactly what it picked. No tool names in the prompt. No hints. Just: what does Claude Code naturally reach for?
The results are worth understanding if you're building anything in 2026.
The Default Stack (JS)
When Claude Code picks a tool, it picks decisively:
- Vercel — 93.8% of frontend deployments (Next.js projects). Zero competitors got a primary pick.
- Stripe — 91.4% for payments
- shadcn/ui — 90.1% for UI components
- Tailwind — 100% of JS frontend styling
- Zustand — 64.8% for state management (while Redux gets zero primary picks, just 23 mentions)
- Sentry — 63.1% for observability
For Python:
- Railway — 82% for deployment (AWS, GCP, Azure got zero primary picks)
- FastAPI — dominant for APIs
- Custom/DIY — increasingly preferred over Redis for caching in newer models
The Most Important Finding: Claude Code Builds, Not Buys
In 12 of 20 categories, Claude Code builds custom solutions rather than recommending any external tool.
Ask it to "add feature flags"? It builds a config system with env vars and percentage-based rollout — not LaunchDarkly.
Ask it to "add auth" in Python? It writes JWT + bcrypt from scratch — not Auth0, not Supabase.
This is the real signal for anyone building a developer tool or SaaS: Claude Code is your competitor, not just a distribution channel. If your product solves something Claude can implement in 50 lines, you have a serious moat problem.
The Recency Gradient: Models Learn Fast
The shift between Sonnet 4.5 and Opus 4.6 is dramatic:
| Category | Sonnet 4.5 | Opus 4.6 |
|---|---|---|
| ORM (JS) | Prisma 79% | Drizzle 100% |
| Jobs (Python) | Celery 100% | FastAPI BgTasks 44% |
| Caching (Python) | Redis 93% | Custom/DIY 50% |
Drizzle went from 21% to 100% between model versions. Celery collapsed from 100% to 0%.
This means: the tools that exist in 2024 are not necessarily the tools that will exist in every AI-generated app by 2027. The model is moving faster than the market.
What This Means If You're Building Something
If you're a founder bootstrapping with AI agents:
You're getting a highly opinionated stack by default. That's mostly good news — Vercel + Railway + Stripe + shadcn is a genuinely solid foundation for a SaaS. But it also means:
- You're probably getting Tailwind even if you didn't ask for it
- Your auth is probably being written from scratch (not a library)
- Your feature flags are env vars, not a proper system
- Your ORM is increasingly Drizzle, not Prisma
Know this. Check your package.json after every major AI session.
If you're building a devtool:
Claude Code is now the gatekeeper. If your tool isn't in the model's training data as a go-to recommendation, you're invisible at the moment of implementation.
The winners (Vercel, Stripe, shadcn) didn't just build good products — they created enough quality documentation, tutorials, and ecosystem presence that the model associates them with the solution instantly.
The losers (Redux, Heroku, AWS Amplify) are mentioned but never chosen. That's a new kind of market irrelevance.
If you're building AI services:
The fact that Claude builds custom solutions means there's a growing gap between "good enough AI-generated code" and "production-hardened, compliant, integrated solutions."
That gap is where services businesses live. An AI can write JWT auth in 50 lines, but it can't:
- Get on your client's approved vendor list
- Handle their HIPAA audit
- Call their office when something breaks
- Set up and monitor the system for 6 months
The services opportunity isn't shrinking because of AI. It's being redefined.
The Tailwind Paradox
One of the more interesting threads in the HN discussion: Tailwind "won" the CSS war in AI models, but the company recently had to let go of staff because revenue has plummeted.
Their docs traffic dropped — developers stopped visiting because they just ask AI. But AI keeps recommending Tailwind. The product is everywhere; the business is struggling.
This is the new paradox for developer tools: ubiquity via AI does not equal revenue. If your monetization depends on documentation traffic, tutorial views, or paid component libraries — AI may recommend your core product while gutting your business model simultaneously.
The Bottom Line
Claude Code has opinions. Strong ones. And because millions of developers are now letting AI write their first draft, those opinions are becoming the defaults of an entire generation of software.
If you're building with AI agents: know what you're getting.
If you're building for developers: start optimizing for model visibility, not just Google.
If you're building services: the custom code Claude writes is your sales pitch, not your competition.
The stack is being decided. Make sure you understand who's deciding it.
Building AI services for businesses? RooxAI helps companies deploy AI that handles real workflows: phone receptionists, customer support, sales automation.
Top comments (0)