inherit-legacy-style

Prevent AI code style drift in legacy projects. A Claude Code skill that scans your hand-written project, extracts its implicit coding conventions, and crystallizes them into enforceable rules — so every subsequent AI edit aligns with the project's existing style, not the model's pretrained biases.

English | 中文

---

English

The Problem

When you drop an AI coding agent onto a hand-written legacy project, the model's pretraining bias makes it impose its own "mainstream" idioms. Result: two competing list-refresh patterns in one codebase, inconsistent naming, duplicated utilities. Style Drift.

What This Skill Does

• Zero-argument trigger: /inherit-legacy-style auto-detects whether it's a first-time scan or an incremental update.
• 4-dimension meta-architecture scan: File Anatomy, State & Control Flow, Infrastructure, Error Handling — language- and framework-agnostic.
• Scale-adaptive: Small projects get full close-read; large projects get smart sampling with a budget cap.
• Signal-threshold noise reduction: Minor inconsistencies are auto-silenced into DONTs; only high-signal conflicts interrupt you for a decision.
• Enforceable: Writes rules to .ai-style-rules.md and hooks them into your project's CLAUDE.md so every future session auto-loads the constraint.

Installation

# Clone into your Claude Code skills directory
git clone https://github.com/wulinzai/inherit-legacy-style.git ~/.claude/skills/inherit-legacy-style

Then reload plugins in Claude Code (/reload-plugins) or restart the session.

Usage

/inherit-legacy-style

That's it. The skill auto-detects whether .ai-style-rules.md exists:

• Not found → Full scan mode: measures project scale, scans along 4 dimensions, resolves conflicts with you, creates the rules file.
• Found → Incremental sniff mode: checks recent Git changes for new patterns not yet recorded, appends evolution log.

What Gets Generated

A root-level .ai-style-rules.md with:

• [Golden Files] — exemplar files to imitate pixel-for-pixel
• [Naming & State-Control Rules] — concrete, checkable conventions
• [DONTs] — legacy anti-patterns that must not propagate

Requirements

• Claude Code (the skill uses Read, Glob, Grep, Bash, Edit, Write tools)
• Git (recommended for incremental mode; non-Git projects fall back to file timestamps)

License

MIT

---

中文

痛点

历史人工项目接入 AI 编程时，大模型会因预训练偏见把自己的"主流写法"强加给项目，造成代码风格漂移（Style Drift）：同一个项目里出现两套列表刷新逻辑、命名割裂、工具函数重复造轮子。

技能能力

• 零参数触发：/inherit-legacy-style 自动检测首次/增量模式。
• 四维元架构扫描：文件内组织、状态控制、基础设施、防御性编程——跨语言、跨技术栈通用。
• 规模自适应：小项目全量精读，大项目智能抽样+预算上限。
• 信号阈值降噪：弱信号自动入 DONTs，仅高信号冲突中断提问。
• 持久约束：规约写入 .ai-style-rules.md 并挂钩到 CLAUDE.md，每会话自动生效。

安装

git clone https://github.com/wulinzai/inherit-legacy-style.git ~/.claude/skills/inherit-legacy-style

在 Claude Code 中 /reload-plugins 或重启会话即可。

使用

/inherit-legacy-style

技能自动检测 .ai-style-rules.md 是否存在：

• 不存在 → 首次全量扫描：量规模、四维扫描、冲突逐个拷问、生成规约文件。
• 已存在 → 增量嗅探：读取 Git 近期改动，比对既有规约，仅追加演进日志。

生成产物

项目根目录 .ai-style-rules.md，包含三大板块：

• [Golden Files] 金牌标杆文件 — 值得像素级模仿的真实文件路径
• [强制命名与状态控制规约] — 可执行、可对照的明确规则
• [DONTs] 禁止模仿的遗留反面教材 — 不应再扩散的坏味道

依赖

• Claude Code（使用 Read、Glob、Grep、Bash、Edit、Write 工具）
• Git（增量模式推荐；非 Git 项目退化为文件时间戳比对）

License

MIT