Vibe Coding、Cursor、Claude Code、Codex、Copilot
This collection features articles and community discussions in the AI 编程 space, focusing on real-world engineering scenarios, tool selection, practical paths, and risk boundaries.
The page is continuously updated and serves as a learning roadmap, technology selection reference, and technical review entry point.
Being able to export chat JSON doesn't mean you can leave a coding tool. This article defines a content-addressed exit package covering repository rules, task checkpoints, approvals, tool configuration, and audit receipts, and uses a cold start with an alternative tool to prove migration and revocation are truly complete.
Example-based tests easily share blind spots with the model implementation. This article extracts invariants from requirements, designs generators, shrinkers, and deterministic seeds, and uses counterexample regression gates to validate the input space of AI-generated code.
Line coverage can't prove a test rejects faulty patches. This article wires mutation testing into the AI change gate, using scoped ranges, surviving-mutant review, and baseline diffs to verify whether tests actually constrain new behavior.
Knowing add/commit/push is not the same as knowing Git. This article explains rebase vs. merge, interactive rebase, cherry-pick, reflog, stash, and practical recovery from common mistakes so Git is no longer intimidating.
The terminal is a developer's home turf, and a well-configured toolchain can double your efficiency. This article explains a modern setup: Zsh + Oh My Zsh/Starship for prompts, Tmux for sessions, panes, and persistence, fzf for fuzzy finding, and modern replacements such as eza, bat, ripgrep, and zoxide.
A codemod can rewrite thousands of call sites at once, and just as reliably replicate the same bug. This article builds an executable invariant gate from node selection, semantic preservation, idempotency, format boundaries, and diff budgets.
A smarter-looking demo of a new code search engine doesn't mean it can replace your existing grep, LSP, or symbol index. This article freezes query tasks, code snapshots, and relevance judgments, then uses dual-index shadow traffic to verify recall, freshness, permissions, and reversible switching.
Claude Code's power isn't 'chatting' — it's that you can customize it into a dedicated tool. Covers two extension mechanisms: Skills (packaging domain knowledge + workflows into reusable capabilities) and Hooks (event-driven automation), with a real example baking workflows like 'run an SEO check before publishing' into Claude Code.
A diff generated by an agent doesn't mean the engineering task is done. This article builds validation gates—static checks, targeted tests, regression scope, migration review, and human confirmation—so every patch carries traceable, mergeable evidence.
A line of text in a terminal doesn't mean a command actually ran in the approved way. This article defines an immutable receipt for command argv, cwd, executable, policy decision, exit status, and output digest, while preventing secrets from spreading through logs.
A larger context window doesn't mean you should hand the whole repo to the model. This article covers task-level token budgets, layered summaries, evidence references, cache invalidation, and compression-quality checks to keep a coding agent's engineering judgment intact within limited context.
A single error message and the final patch aren't enough to prove a bug was reproduced. This article defines a content-addressed debug evidence bundle that freezes code, uncommitted state, dependencies, inputs, and failure criteria, so an isolated runner can first observe the old failure, then verify the new fix.