Tried to replicate the Cursor official website and felt frustrated, so I turned back to look at Agent again.
https://01.me/2025/12/silicon-valley-ai-insights-2025/
In the section on AI coding effectiveness, the day-to-day development at big tech companies struck me as funny — the time spent actually writing code is very limited, only 15%. I really don't like that.
Research Code can be used to write agents, write scripts, and so on.
Infrastructure code, including the Linux kernel and consensus protocols, doesn't work so well.
The best practice for vibe coding is to split: generate as little code as possible at a time.
Another one is TDD, test-driven development. Honestly, I think it's a bit more reliable than Ralph-style development...
Specs are important for large-scale refactoring. I even saw a paper about writing a Linux kernel file system from a spec... not sure how well it works.
A strict evaluation system is also a process of accumulating code data. Everyone now knows data matters; every company is building datasets.
The stuff about the Silicon Valley giants was really eye-opening.
The insights about startups are also helpful, I think. In fact, a startup needs to find its own niche in the industry — it can't do general-purpose areas because big companies all do those; it has to find a very specific vertical.
I think one must never detach from engineering practice. Only by trying things with your own hands can you get the truest feel for a piece of work — whether it's vibe coding or training models. You can't rely on hearsay; you have to try it yourself.
Technical Practice
Context Engineering Framework
- System Prompt
- Tools
- Data Retrieval
- Long Horizon Optimizations
The Data-Retrieval paradigm shift
The new approach is just-in-time loading
- Strategy one: lightweight identifiers
- Progressive disclosure
- Autonomous exploration
All models show performance degradation on long contexts
Solutions when the context window capacity is exceeded
- Context compression
- The Agent maintains explicit memory artifacts storing "work notes": decisions, learnings, state. Retrieved on demand, rather than kept in context
- Sub-Agents. Decompose complex tasks into specialized Agents; each Sub-Agent has a focused, clear, narrow context. The main agent orchestrates and synthesizes the results
How the Skills mechanism works
Claude can dynamically discover and load them
pdf/SKILL.md (主文件)
├── YAML Frontmatter (name, description)
├── Overview (概述)
└── References: "For advanced features, see /reference.md"
pdf/reference.md (详细参考)
└── Advanced PDF processing features...
pdf/forms.md (专门功能)
└── PDF form filling instructions...- Memory
- Sub Agents & Collaboration
- Dynamic Tool Calls
- Code Generation & Execution
- Web Search
- Agentic Search
- Long Running Tasks