PluginBench
Skill
Pass
Audit score 90

improve-codebase-architecture

vinvcn/mattpocock-skills-zh-cn

Scan codebases for architectural deepening opportunities and generate visual refactoring reports.

What is improve-codebase-architecture?

Identifies shallow modules and architectural friction points in your codebase, generates an interactive HTML report with before/after visualizations, then guides you through deepening decisions. Use this when you want to improve testability, reduce cognitive load, and make future changes easier in actively-changing code areas.

  • Analyzes commit history to find hot-spot areas that change frequently
  • Applies deletion test heuristics to identify shallow modules worth deepening
  • Generates visual HTML reports with Mermaid diagrams showing current vs. improved architecture
  • Provides structured grilling loop to walk through refactoring constraints and dependencies
  • Updates domain model (CONTEXT.md) and proposes ADRs for load-bearing architectural decisions

How to install improve-codebase-architecture

npx skills add https://github.com/vinvcn/mattpocock-skills-zh-cn --skill improve-codebase-architecture
Prerequisites
  • Project with git history (to identify hot-spot areas)
  • Optional: CONTEXT.md file with domain vocabulary
  • Optional: docs/adr/ directory with existing architectural decision records
Claude Code
Cursor
Windsurf
Cline

How to use improve-codebase-architecture

  1. 1.Run the skill and optionally specify a module, subsystem, or pain point to focus on
  2. 2.If no direction given, the skill analyzes recent commit history to find frequently-changing areas
  3. 3.Review the generated HTML report (opens automatically) showing candidate refactorings with before/after diagrams
  4. 4.Select a candidate from the report to explore further
  5. 5.Work through the grilling loop to finalize constraints, seams, and deepened module shape
  6. 6.Domain model and ADRs update automatically as decisions are made

Use cases

Good for
  • Refactoring a frequently-modified subsystem to reduce cognitive load when making changes
  • Improving testability by consolidating scattered pure functions into deep modules with clear seams
  • Identifying tightly-coupled modules that leak across architectural boundaries
  • Planning a multi-step deepening strategy for a complex domain model
  • Recording architectural decisions to prevent future reviews from re-suggesting rejected refactors
Who it's for
  • Backend and full-stack engineers managing growing codebases
  • Architects planning refactoring campaigns in active projects
  • Teams using domain-driven design and architectural decision records
  • Developers working with Claude Code or Cursor agents on codebase improvements

improve-codebase-architecture FAQ

What is a 'deep' vs 'shallow' module?

A deep module has a simple, powerful interface that hides substantial implementation complexity. A shallow module has an interface nearly as complex as its implementation, forcing callers to understand internal details. Deep modules reduce cognitive load and improve testability.

How does the deletion test work?

Imagine deleting a module entirely. If complexity concentrates in fewer places (good signal), the module is worth deepening. If complexity just moves elsewhere without concentrating, the current structure may already be reasonable.

What if a candidate conflicts with an existing ADR?

The skill only flags conflicts when friction is real enough to justify reopening the ADR. Theoretical conflicts are ignored. You can propose recording a new ADR if a candidate is rejected for load-bearing reasons.

Do I need CONTEXT.md and ADRs to use this skill?

No, they are optional. The skill works without them but uses them when present to ground suggestions in your domain vocabulary and avoid re-suggesting rejected decisions.

Can I focus the scan on a specific area instead of hot-spot detection?

Yes. Name the module, subsystem, or pain point when running the skill, and it will skip commit-history analysis and explore only that area.

Full instructions (SKILL.md)

Source of truth, from vinvcn/mattpocock-skills-zh-cn.


name: improve-codebase-architecture description: 扫描代码库中的深化机会,生成可视化 HTML 报告,然后围绕你选中的候选项继续追问。 disable-model-invocation: true

Improve Codebase Architecture

暴露 architectural friction,并提出 deepening opportunities:把 shallow modules 变成 deep modules 的 refactors。目标是 testability 和 AI-navigability。

这个命令由项目的 domain model 提供信息,并建立在共享 design vocabulary 上:

  • 调用 Skill 工具并指定 codebase-design,获取 architecture vocabulary(module、interface、depth、seam、adapter、leverage、locality)及其 principles(deletion test、"the interface is the test surface"、"one adapter = hypothetical seam, two = real")。每条建议都准确使用这些术语,不要漂移到 "component"、"service"、"API" 或 "boundary"。
  • CONTEXT.md 中的 domain language 会为好的 seams 命名;docs/adr/ 中的 ADRs 记录这个命令不应重新争论的 decisions。

Process

1. Explore

先划定扫描范围——YAGNI。 深化 module 的收益在于让未来修改更容易,因此要更关注最近仍在变化的 codebase 区域。开始探索前先决定去哪里看:

  • 如果用户点名了方向——module、subsystem 或 pain point——就按该方向探索,跳过下面的推断。
  • 否则,向前回看一段足够长的 commit history(git log --oneline),找出反复出现的 files 和 areas,让这些 hot spots 成为首要关注点。如果变更分散、没有明显 hot spot,再扩大范围。

先读取项目 domain glossary(CONTEXT.md)以及你将触碰区域的 ADRs。

然后 spawn 一个 sub-agent 来遍历 codebase。不要套死板 heuristics;自然探索,并记录你感到 friction 的地方:

  • 理解一个概念是否需要在许多小 modules 之间来回跳?
  • 哪些 modules 是 shallow 的,即 interface 几乎和 implementation 一样复杂?
  • 是否存在为了 testability 抽出的 pure functions,但真正 bugs 藏在它们如何被调用之处(没有 locality)?
  • 哪些 tightly-coupled modules 泄漏到了 seams 之外?
  • Codebase 的哪些部分未测试,或很难通过当前 interface 测试?

对任何你怀疑 shallow 的东西应用 deletion test:删除它会让复杂度集中,还是只把复杂度移动到别处?"yes, concentrates" 才是你要的 signal。

2. Present candidates as an HTML report

把 self-contained HTML file 写到 OS temp directory,避免任何内容落进 repo。Temp dir 从 $TMPDIR 解析,fallback 到 /tmp(Windows 用 %TEMP%),写到 <tmpdir>/architecture-review-<timestamp>.html,让每次运行都有新文件。为用户打开它:Linux 用 xdg-open <path>,macOS 用 open <path>,Windows 用 start <path>,并告诉用户 absolute path。

Report 使用 Tailwind via CDN 做 layout/styling,用 Mermaid via CDN 做能可靠传达结构的 diagrams。Mermaid 和手写 CSS/SVG visuals 可以混用:关系是 graph-shaped(call graphs、dependencies、sequences)时用 Mermaid;需要 editorial 表达(mass diagrams、cross-sections、collapse animations)时用手写 divs/SVG。每个 candidate 都要有 before/after visualisation。要视觉化。

每个 candidate 渲染一张 card,包含:

  • Files - 涉及哪些 files/modules
  • Problem - 当前 architecture 为什么造成 friction
  • Solution - 会改变什么,用平实的语言描述
  • Benefits - 用 locality 与 leverage 解释收益,以及 tests 如何改善
  • Before / After diagram - side-by-side,自绘,说明 shallowness 与 deepening
  • Recommendation strength - Strong、Worth exploring、Speculative 之一,渲染为 badge

Report 末尾包含 Top recommendation section:你会先处理哪个 candidate,以及为什么。

用 CONTEXT.md vocabulary 表达 domain,用 /codebase-design vocabulary 表达 architecture。 如果 CONTEXT.md 定义了 "Order",就说 "Order intake module",不要说 "FooBarHandler",也不要说 "Order service"。

ADR conflicts:如果 candidate 与现有 ADR 冲突,只有在 friction 真实到值得重新打开 ADR 时才提出。Card 中明确标记(例如 warning callout:"contradicts ADR-0007 - but worth reopening because...")。不要列出 ADR 理论上禁止的每个 refactor。

完整 HTML scaffold、diagram patterns 和 styling guidance 见 HTML-REPORT.md。

现在不要提出 interfaces。写完文件后问用户:"Which of these would you like to explore?"

3. Grilling loop

用户选中 candidate 后,调用 Skill 工具并指定 grilling,与用户走完 decision tree:constraints、dependencies、deepened module 的形状、seam 后面放什么、哪些 tests 能保留。

Side effects 随 decisions 成形而内联发生;调用 Skill 工具并指定 domain-modeling,让 domain model 保持最新:

  • 要用 CONTEXT.md 中不存在的概念命名 deepened module? 把 term 加入 CONTEXT.md。若文件不存在,按需创建。
  • 对话中打磨了 fuzzy term? 立即更新 CONTEXT.md。
  • 用户以 load-bearing reason 拒绝了 candidate? 提议写 ADR,表述为:"Want me to record this as an ADR so future architecture reviews don't re-suggest it?" 只有当未来 explorer 确实需要该 reason 以避免再次提出同样建议时才提议;跳过临时原因("not worth it right now")和显而易见原因。
  • 想探索 deepened module 的 alternative interfaces? 调用 Skill 工具并指定 codebase-design,并使用其中的 design-it-twice parallel sub-agent pattern。