Kuangjia Chaijie - Decomposição Estruturada de Frameworks
eiaao/kuangjia-chaijie
Decomponha frameworks completamente antes de escrever sobre eles
Ver no GitHubSkill Claude Code que realiza análise profunda e sistemática de frameworks/codebases através de pipeline agentic com 4-5 fases estruturadas. Combina varredura top-down (modelo mental) com bottom-up (rastreamento de símbolos) e checkpoints humanos obrigatórios para garantir compreensão completa e sem lacunas antes de escrever artigos técnicos ou apresentações.
Instalação
npx convez-ai eiaao/kuangjia-chaijie1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Preparar artigos técnicos e palestras com compreensão profunda
- Documentar arquitetura de sistemas complexos multi-módulo
- Análise de codebase para onboarding ou auditoria técnica
Veredito

Skill bem-pensado para quem precisa realmente entender sistemas complexos antes de documentá-los ou ensinar. Pipeline rigoroso com human-in-the-loop evita o típico problema de análise superficial; excelente para arquitetos e technical writers, mas demanda disciplina e tempo. Ótimo para empresas que valorizam documentação profunda.
README
A tradução para o português deste README ainda está na fila.
kuangjia-chaijie · Framework Topic Decomposition Workflow
Understand a framework deeply — completely and gap-free — before you write or present about it.
English | 中文
This skill does not output an article. It outputs "complete, out-loud-recountable understanding." Input: a framework/codebase + a topic word → Output: 4 (+1 optional) structured markdown documents. Writing the article is what happens after the skill finishes.
For anyone preparing a tech talk, a deep-dive article, or a team session: before you start writing, get a cross-module topic complete, clear, and gap-free in your head first.
The problem it solves
Three failure modes dominate when reading framework source:
- Guessing — filling gaps with "usually / typically / by convention" from training-data patterns, which often differ from how this project is actually implemented.
- Missing — the topic spans modules; a top-down scan based on current understanding can't catch modules hidden mid-file in a large file, or on a background path off the main loop.
- Drifting — shallow reads produce shallow conclusions, and self-review skips over what you already wrote, so errors survive to the end.
This skill counters them with an agentic pipeline: top-down (project mental-model scan) + bottom-up (symbol back-tracing) in tandem, mandatory human-in-the-loop checkpoints for scope, hard exit criteria per phase, and cross-verification of every phase's output by a clean-context subagent against the source.
When to use / not use
✅ Use it when
- Preparing a tech talk / article / team session but not yet on top of the topic
- The topic spans multiple modules (not explainable from a single file)
- You want "complete, no gaps" understanding of a module
❌ Don't use it for
- Single-file / single-function questions → just Read / Grep
- Debugging a specific bug → use a debugging skill
- Designing a new feature → use brainstorming → writing-plans
Pipeline (4 phases + 1 optional)
| Phase | What | Output | Exit gate (excerpt) |
|---|---|---|---|
| 0 Seeding | Turn a fuzzy topic word into a "seed card" | 00-seed.md | ≥5 keywords, ≥2 explicit exclusions |
| 1 Panorama (A+B) | Mental model + symbol heatmap + cross-check + panorama + reverse coverage + blind-spot review | 01-*.md | 5–9 modules, ≥2 Mermaid diagrams, subagent-verified, user sign-off |
| 2 Per-module deep dive | A "7-element card + design trade-offs + retro" per module | 02-S{n}-{slug}.md (one each) | All 10 blocks filled, lifecycle diagram, subagent-verified |
| 3 Re-unification | v0→v1, architecture diagram, cross-module patterns, cross-card conflicts, upstream backfill | 03-synthesis.md | One-sentence essence, ≥2 cross-module patterns, conflicts explicitly resolved, backfill table processed |
| 4 Evolution archaeology (optional) | Pick modules by interest; funnel-style git archaeology of "how it became what it is" | 04-evolution-S{n}-{slug}.md | Timeline diagram + inflection points with commit hashes + why, stability annotations |
Four core principles
- Don't guess → Read first → then ask. Every conclusion needs evidence (code/docs/data you read, or what the user explicitly told you). Words like "usually / generally / should / likely" trigger a stop: either go Read for evidence or mark
[to clarify: ...]. - Visualization is mandatory. Plain text + tables don't count; each phase needs at least one Mermaid diagram — drawing is the most efficient blind-spot finder, and a connection you can't draw is a real gap.
- Clean-context subagent verification. Each phase's output goes to a subagent that has "never seen this card" to check claims against source line by line. Same context = same blind spots; main-agent self-verification is no verification.
- A feedback loop, not a one-way pipe. After each phase, revisit upstream outputs; apply the Amend rules for any errors found (small fixes in place, structural revisions logged in amendments, abstraction shifts via v0→v1).
Install
Option 1: Plugin marketplace (recommended)
In Claude Code, run:
/plugin marketplace add eiaao/kuangjia-chaijie
Then Browse and install plugins → kuangjia-chaijie → Install.
Option 2: Manual
Copy the skill directory into Claude Code's skills folder:
git clone https://github.com/eiaao/kuangjia-chaijie.git
cp -r kuangjia-chaijie/skills/kuangjia-chaijie ~/.claude/skills/
Usage
Just tell Claude your topic, e.g.:
- "I want to deep-dive the memory architecture of the Hermes framework"
- "Help me get clear on LangChain's tool-calling mechanism"
- "I'm writing an article about part of framework XX — sort it out for me first"
- "Give me the panorama of XX first"
The skill starts at Phase 0 and pauses at checkpoints for you to confirm scope. Outputs land by default in notes/kuangjia-chaijie/{topic-slug}/ under the project root.
Layout
kuangjia-chaijie/
├── .claude-plugin/marketplace.json # plugin marketplace manifest
├── skills/kuangjia-chaijie/
│ ├── SKILL.md # the workflow itself
│ └── assets/
│ ├── build_view.py # renders output .md into a single-page HTML view
│ └── template.html # HTML view template
├── README.md # English (this file)
└── README.zh.md # Chinese
assets/build_view.py aggregates each phase's markdown outputs into an interactive single-page view.html (markdown stays the source of truth; the HTML is just a view layer, gitignored).