谈谈 AI 编程工具的进化与 Vibe Coding

“谈谈 AI 编程工具的进化与 Vibe Coding” (roughly, “On the evolution of AI programming tools, and vibe coding”) is a personal blog post published on guangzhengli.com on 28 August 2025. Its author, who says they began using GitHub Copilot in 2023, used Cursor continuously from 2024 and had been using Claude Code heavily for a few months before writing, argues that Vibe Coding has become a label for every way of writing code with AI, and that much of the current argument about it comes from not separating two different practices.

The post keeps “vibe coding” for the sense of the Andrej Karpathy post it quotes, which it reads as four points: forgetting that the code exists; fixing even small errors through the AI rather than by hand; no longer reviewing what the AI writes and judging only the result; and accepting that this is not bad, and fairly fun, for throwaway projects. For AI-assisted programming in general it proposes the name Context Coding, and spends most of its length on that practice, reading the history of the best-known coding tools as a history of Context engineering.

Key Points

• The author proposes “Context Coding” as a better name for AI-assisted programming than vibe coding, on the view that, with the model held fixed, every improvement in AI-assisted programming rests on passing the LLM more suitable context — whether through chat, RAG, rules or MCP.
• GitHub Copilot’s early success is attributed to being the first tool to bring code context together for the LLM — the code in the open editor window and the code around the cursor. Its early limits, as the author recalls them, were a model that still hallucinated heavily on code (GPT-3.5 at the time), a small context, no ability to edit code directly, and no view of any file other than the one open.
• Cursor is credited with indexing a project’s entire codebase for semantic (vector) search through RAG, and with letting users attach files and folders to the context themselves; the author reads this fuller, more user-controlled context as why it overtook Copilot.
• Claude Code is described as feeding the model context generously: it first analyses a codebase’s structure and technology stack, then retrieves code with Unix tools such as grep, find, git and cat rather than RAG. In the author’s experience the two tools are similar on small and medium tasks, while on large tasks touching more than ten files Claude Code does far better, which the author attributes to Anthropic, as the model provider, being less constrained in how many tokens it spends.
• The author treats both sides of the RAG-versus-grep dispute as having a point, predicts that full AI IDEs will come to offer both kinds of retrieval, and argues that terminal tools such as Claude Code and Gemini CLI need not adopt RAG, because their strength also lies in scriptable workflows that integrate with codebases, MCP and CI/CD pipelines.
• Practical advice, drawn from the author’s own practice: give the model what a new team member would need — the technology stack, the directory structure, what files are named and for what, common commands such as install, lint, test and build — in an instruction file such as GitHub Copilot’s .github/copilot-instructions.md or Claude Code’s CLAUDE.md. The author warns that such context is not better the more there is of it: information that goes stale without the file being updated, such as directory layouts or frequently refactored utilities, does more harm than leaving it out.
• For complex tasks the author suggests having the model split the work into subtasks recorded in a document, update that document as each one is finished, commit in small steps and delete the document at the end, and reports that this markedly reduces hallucination.
• Other sources of context the post recommends are current documentation passed in through MCP servers such as context7, and having the model add logging throughout faulty code as a stand-in for a debugger’s view of inputs and outputs. It singles out Claude Code’s built-in /context command, which shows how much of the context window each kind of content occupies and how much remains, as the first tool the author had seen expose context usage to the user.
• The author’s own practice is mixed rather than agent-only: in serious engineering work, Cursor’s Tab completion is what the author uses most, writing the layering and abstractions by hand and letting completion fill in the code, because in the author’s experience LLMs remain poor at abstraction.
• On vibe coding proper, the post argues that for people without programming experience it introduces defects and security vulnerabilities in the short term and unmaintainable code and technical debt in the long term. It borrows the comparison of handing a child a credit card before explaining debt, and recounts an X user whose product, built entirely by vibe coding, was shut down days after the post announcing it drew attention: the product was attacked, its API key usage hit the maximum, and someone bypassed its subscription.
• On careers, the author states a pessimistic view: that most programmers work as translators of natural-language requirements into code, that LLMs can increasingly fill that role, and that the number of average programmers will shrink, while AI’s leverage will make independent developers and small teams more common.

Context

The post is one practitioner’s opinion, grounded in the author’s own use of the three tools it discusses rather than in any measurement, and the author describes it as written in haste and full of personal views. Its account of how Cursor’s indexing works is a brief explanation in the author’s words; the post’s reference list includes Cursor’s security page on codebase indexing. On naming, the post offers Context Coding as one of several names it considers better than vibe coding for AI-assisted programming — alongside AI-assisted programming, AI Coding and Agents Coding — and states a personal preference for it. For another piece of writing on where the boundary of vibe coding lies, see Not all AI-assisted programming is vibe coding (but vibe coding rocks).