Trying to recreate the Cursor homepage left me frustrated, so I pivoted back to agents.
https://01.me/2025/12/silicon-valley-ai-insights-2025/

The “big-company day-to-day dev” bit under AI coding effects is almost funny—so little time actually writing code (~15%). Not how I want to work.

Research code is fine for agents and scripts.

Infrastructure code—Linux kernel, consensus protocols—not really there yet.

Vibe coding best practice: split work; generate as little code per step as possible.

TDD honestly feels more reliable than “Ralph-style” dev to me…

Large refactors need a solid spec—I’ve even seen papers that write a Linux filesystem from spec… curious how well that works.

A strict eval system is also a way to accumulate code data—everyone knows data matters now; each company builds datasets.

The giant Silicon Valley dynamics were eye-opening.

Startup takeaways were useful too—you need a real niche; don’t compete in generic lanes the big players will own. Find a narrow vertical.

You can’t skip engineering—only by doing it do you get a true feel for the work. Vibe coding, training models, whatever: don’t trust hearsay; try it yourself.

Technical practice

Context engineering stack

• System prompt
• Tools
• Data retrieval
• Long-horizon optimizations

Data retrieval paradigm shift: the new pattern is just-in-time loading:

• Strategy 1: lightweight identifiers
• Progressive disclosure
• Autonomous exploration

All models degrade on long context. When you exceed the window:

• Context compression
• Agents maintain explicit memory artifacts—working notes of decisions, learning, state—retrieved on demand instead of stuffed into context
• Sub-agents: decompose into specialists with narrow, clear context; the main agent orchestrates and synthesizes

How skills work

Claude can discover and load them dynamically.

pdf/SKILL.md (main file)
├── YAML Frontmatter (name, description)
├── Overview
└── References: "For advanced features, see /reference.md"

pdf/reference.md (deep reference)
└── Advanced PDF processing features...

pdf/forms.md (specialized)
└── PDF form filling instructions...

• Memory
• Sub-agents & collaboration
• Dynamic tool calls
• Code generation & execution
• Web search
• Agentic search
• Long-running tasks