Big Tech's massive borrowing spree

Wait, the title is about "Big Tech's massive borrowing spree," but we're in the "AI Programming" section? Let's connect the dots: the reason these giants are piling on debt is to fund the insane capex for H100 clusters and custom silicon. For us devs, this "borrowing spree" translates directly into the massive LLM context windows and the "Compute-Optimal" models we're using in Cursor and Claude Code right now.

I’ve been spending the last month trying to squeeze every drop of productivity out of Claude 3.5 Sonnet via Cursor, and it's clear that the sheer scale of the infrastructure these companies are buying is fundamentally changing how I architect code. I'm no longer writing "defensive" boilerplate to save on token costs or avoid hallucination; I'm shifting toward a "Context-First" workflow.

The biggest productivity jump for me hasn't been the autocomplete, but the .cursorrules file. If you aren't using a project-specific rules file, you're basically leaving 30% of the AI's capability on the table. Instead of repeating "use Tailwind" or "follow the Repository pattern" in every prompt, I dump my entire architectural philosophy into a hidden file.

Here is the exact snippet I use to stop the AI from suggesting outdated libraries or weird patterns:

# Project Standards
- Use TypeScript strict mode.
- Prefer functional components over classes.
- State management: Use Zustand, avoid Redux unless specifically asked.
- Styling: Tailwind CSS only. No CSS-in-JS.
- Error Handling: Wrap all async calls in a try-catch block with a standardized logger.
- API: Use Zod for schema validation on all incoming request bodies.

Once this is in place, the "Composer" mode (Cmd+I) becomes a beast. I can literally say "Add a user profile page with an upload button," and it doesn't just write the JSX—it knows to use Zod for the validation and Zustand for the state because it's reading the rules.

One major gotcha I've hit: the "Context Drift." When you let an AI agent like Claude Code or Cursor's Composer edit five files at once, it occasionally forgets a constraint from the first file by the time it hits the fifth. I've found that the only way to kill this is to force a "Verification Step."

I now use a specific prompt pattern to audit the changes before I commit:

Review the changes you just made across all files. 
Check for:
1. Any unused imports introduced.
2. Type mismatches between the API response and the frontend interface.
3. Consistency with the .cursorrules file.
If you find an error, fix it now.

The productivity gain is tangible. I'm moving from "writing code" to "reviewing diffs." I estimate my velocity has increased by about 2x, but the cognitive load has shifted. I spend less time worrying about syntax and more time worrying about system design. If Big Tech is borrowing billions to build these clusters, the least we can do is stop writing manual boilerplate and start acting like architects.