38 Skills + 4 Agents + 12 Rules + 7 Hooks — forged from 2,137 community assets
38 个 Skills + 4 个 Agents + 12 条 Rules + 7 个 Hooks — 从 2,137 条社区资产中基因锻造而来
EN: A production-ready Claude Code configuration — not a collection of copy-pasted community files, but a gene-forged arsenal. Every weapon was created by scanning 2,137 community assets across 23 high-star repos, dissecting them into atomic elements, extracting only expert-level knowledge Claude doesn't already have, and forging it into concise, actionable Skills/Agents/Rules/Hooks.
CN: 一套生产级 Claude Code 配置 — 不是社区文件的复制粘贴合集，而是一个基因锻造的武器库。每件武器都经过：扫描 23 个高星仓库的 2,137 条社区资产 → 元素解剖 → 只提取 Claude 不具备的专家级知识 → 锻造为精炼的 Skills/Agents/Rules/Hooks。
Community Asset (e.g. 14,813 chars)
↓ Element Dissection / 元素解剖
Expert Knowledge + Activation Hints + Redundant Content
↓ Filter / 过滤
Expert Only (Claude doesn't know this)
↓ Forge / 锻造
Concise Skill (~150 lines, pure knowledge delta)
Core formula / 核心公式:
Good Skill = Expert-only Knowledge − What Claude Already Knows
| Category | Skill | Description |
|---|---|---|
| Search & Retrieval | hunt |
4-step progressive search + anti-bias checks |
deep-research |
5-stage pipeline + parallel search + cross-validation | |
rag |
Orama hybrid search + 19-language text splitting | |
distillery-search |
BM25+vector search over 2,137 community assets | |
retrieval-methodology |
Intent classification + channel routing + search strategy | |
context7 |
Live documentation retrieval via Context7 API | |
| Engineering | tdd-workflow |
Red-green-refactor decision trees + Playwright flaky isolation + EDD |
api-design |
URL decision table + HTTP method matrix + pagination + rate limiting | |
docker-patterns |
Multi-stage Dockerfile + security hardening + network isolation | |
database-expertise |
Index selection matrix (B-tree/GIN/BRIN/Hash) + RLS + concurrency | |
code-refactoring |
knip/depcheck + SAFE/CAREFUL/RISKY grading + debt budget | |
mcp-development |
@modelcontextprotocol/sdk stdio server development guide | |
office-documents |
docx/xlsx/pdf/pptx creation and editing | |
| Methodology | context-engineering |
5 degradation patterns + compression strategies + tokens-per-task |
prompt-engineering |
7-framework selection matrix + Claude XML optimization | |
tool-design |
Consolidation principle + description engineering + MCP naming | |
parallel-debugging |
ACH multi-hypothesis + evidence strength grading | |
multi-agent-team |
Parallel dispatch + file ownership + 5-stage pipeline | |
| Meta (Self-Evolution) | skill-creator |
Skill creation wizard |
writing-skills |
TDD approach to writing Skills | |
skill-judge |
8-dimension evaluation (120 points) + element dissection forge | |
skill-testing |
TDD for Skills: 7 pressure types + rationalization capture | |
self-learning |
Autonomous web-based skill generator | |
knowledge-evolution |
Hooks-driven instinct learning with confidence decay | |
find-skills |
Discover and install community skills | |
spiral-bootstrap |
Double-helix evolution orchestrator with L1-L4 graded response | |
evolution-session |
Independent evolution session executor for L3 tool forging | |
| Automation | ralph-loop |
Stop Hook self-loop for unattended iteration with completion signals |
hookify |
Declarative Hook rules via Markdown+YAML with 6 operators | |
| Code Review | pr-review-toolkit |
6-perspective parallel PR review (code/test/error/type/comment/simplify) |
code-review |
5-agent compliance review with per-issue confidence scoring | |
| Workflow | feature-dev |
7-phase gated pipeline: Discovery→Explore→Questions→Design→Implement→Review→Summary |
| Security | security-research |
7-stage penetration testing workflow + layered loading |
exploit-feasibility |
Binary exploitability assessment + mitigation analysis | |
fuzzing-automation |
AFL++ integration + crash collection + corpus management | |
binary-analysis |
Crash analysis + GDB/LLDB debugging + reverse engineering | |
| Quality | requesting-code-review |
Size triage + content-aware routing + static analysis pre-check |
agent-browser |
Browser automation CLI for web interaction |
| 分类 | Skill | 说明 |
|---|---|---|
| 搜索检索 | hunt |
4 步渐进搜索 + 反偏见检查 |
deep-research |
5 阶段 Pipeline + 并行搜索 + 交叉验证 | |
rag |
Orama 混合搜索 + 19 语言分块 | |
distillery-search |
2,137 条社区资产 BM25+向量搜索 | |
retrieval-methodology |
意图分类 + 通道选择 + 搜索策略 | |
context7 |
通过 Context7 API 实时查文档 | |
| 工程实践 | tdd-workflow |
红绿循环决策树 + Playwright flaky 隔离 + EDD |
api-design |
URL 决策表 + HTTP 方法矩阵 + 分页 + 限流 | |
docker-patterns |
多阶段 Dockerfile + 安全加固 + 网络隔离 | |
database-expertise |
索引选择矩阵 + RLS 优化 + 并发模式 | |
code-refactoring |
knip/depcheck 工具链 + 风险分级 + 债务预算 | |
mcp-development |
基于 @modelcontextprotocol/sdk 的 MCP Server 开发 | |
office-documents |
docx/xlsx/pdf/pptx 生成与编辑 | |
| 方法论 | context-engineering |
5 种退化模式 + 压缩策略 + tokens-per-task |
prompt-engineering |
7 框架选择矩阵 + Claude XML 优化 | |
tool-design |
合并原则 + Description Engineering + MCP 命名 | |
parallel-debugging |
ACH 多假设竞争 + 证据强度分级 | |
multi-agent-team |
并行派发 + 文件所有权 + 5 阶段 pipeline | |
| 元能力（自进化） | skill-creator |
Skill 创建向导 |
writing-skills |
TDD 方式写 Skill | |
skill-judge |
8 维度评分（120 分制）+ 元素解剖锻造 | |
skill-testing |
TDD for Skills：7 种压力类型 + 合理化捕获 | |
self-learning |
自主 Web 搜索生成 Skill | |
knowledge-evolution |
Hooks 驱动的本能学习 + 置信度衰减 | |
find-skills |
搜索安装社区 Skills | |
spiral-bootstrap |
双螺旋进化编排器 + L1-L4 分级响应 | |
evolution-session |
独立进化夜班执行器 + L3 工具锻造 | |
| 自动化 | ralph-loop |
Stop Hook 自循环 + 完成信号匹配 + 无人值守迭代 |
hookify |
Markdown+YAML 声明式 Hook 规则 + 6 种 operator | |
| 代码审查 | pr-review-toolkit |
6 视角并行 PR 审查 (code/test/error/type/comment/simplify) |
code-review |
5-Agent 合规性审查 + 逐 issue 置信度评分 | |
| 工作流 | feature-dev |
7 阶段硬门控 Pipeline：Discovery→Summary + 三哲学架构对比 |
| 安全研究 | security-research |
7 阶段渗透测试工作流 + 分层加载 |
exploit-feasibility |
二进制漏洞可利用性评估 + 缓解措施分析 | |
fuzzing-automation |
AFL++ 集成 + 崩溃收集 + 语料库管理 | |
binary-analysis |
崩溃分析 + GDB/LLDB 调试 + 逆向工程 | |
| 质量守护 | requesting-code-review |
大小分级 + 内容感知路由 + 静态分析预检 |
agent-browser |
浏览器自动化 CLI |
| Agent | Role | Tools | Description |
|---|---|---|---|
architect |
Architecture Design / 架构设计 | Read-only | 4-phase workflow + ADR template + 8 anti-pattern detection |
security-reviewer |
Security Audit / 安全深审 | Read-only | OWASP deep review + language-specific patterns (JS/Py/Go) + severity SLA |
retrieval-orchestrator |
Search Dispatch / 检索调度 | All search tools | Intent decomposition + multi-channel dispatch + cross-validation |
memory-updater |
Memory Maintenance / 记忆维护 | Read, Edit, Glob, Grep | Analyze dirty files and update MEMORY.md |
| Rule | Purpose / 用途 |
|---|---|
coding |
5 golden rules + TDD + context efficiency + 3-tier architecture |
error-discipline |
3-Strike escalation + error persistence + anti-repeat failure |
confidence-gate |
3-tier confidence gate (>=90% act, 70-89% options, <70% ask) |
security-gate |
Pre-commit security checklist + secret management |
code-review-standards |
4-level severity + structured output template |
git-workflow |
Conventional Commits + atomic split + PR workflow |
dependency-management |
Pre-add checklist + license compliance + audit |
naming |
Multi-language naming conventions (JS/TS/Python/DB) |
documentation |
What to / NOT to document |
lead-discipline |
Lead role constraints: delegate, don't execute |
handoff-protocol |
Agent handoff document format + verdict system |
language |
Chinese discussion, English code/config |
| Hook | Trigger | Purpose / 用途 |
|---|---|---|
pre-compact-save |
PreCompact | Backup MEMORY.md before context compaction |
post-compact-restore |
SessionStart | Inject MEMORY.md into context after compaction |
stop-guard |
Stop | Prevent accidental stops during autonomous iteration |
post-tool-dirty-tracker |
PostToolUse | Track files modified by Edit/Write/Bash |
pre-tool-guard |
PreToolUse | Hard-block writes to mcp-servers/, warn on main branch |
prompt-submit |
UserPromptSubmit | Auto-read MEMORY.md on first prompt per session |
pain-tracker |
PostToolUse | Sliding-window detection of repeated searches → pain ledger |
CLAUDE.md — System BIOS (bootloader, immutable rules, <=50 lines)
MEMORY.md — State Machine (breakpoint resume, overwrite-only, <=30 lines)
ARMORY.md — Arsenal Registry (active weapons, retired list, vacant niches)
┌─────────────────────────────────────────────────┐
│ Layer 6: Governance / 治理层 │
│ 12 Rules + 7 Hooks │
├─────────────────────────────────────────────────┤
│ Layer 5: Quality / 质量守护层 │
│ security-reviewer, requesting-code-review, │
│ security-gate, code-review-standards │
├─────────────────────────────────────────────────┤
│ Layer 4: Methodology / 方法论层 │
│ context-engineering, prompt-engineering, │
│ tool-design, multi-agent-team, architect │
├─────────────────────────────────────────────────┤
│ Layer 3: Engineering / 工程实践层 │
│ tdd-workflow, api-design, docker-patterns, │
│ database-expertise, code-refactoring, │
│ mcp-development, office-documents │
├─────────────────────────────────────────────────┤
│ Layer 2: Retrieval / 检索层 │
│ hunt, rag, deep-research, distillery-search, │
│ context7, retrieval-methodology, │
│ retrieval-orchestrator │
├─────────────────────────────────────────────────┤
│ Layer 1: Meta (Self-Evolution) / 元能力层 │
│ skill-creator, skill-judge, skill-testing, │
│ self-learning, knowledge-evolution, find-skills │
└─────────────────────────────────────────────────┘
Pain Signal (repeated search, empty results, stale cache)
│
▼
┌─────────────────────────────────────────────────┐
│ spiral-bootstrap (L1-L4 Graded Response) │
│ │
│ L1: Silent log (0 cost) │
│ L2: Data distillation (~2K tokens) │
│ L3: Tool forging (~50K tokens, async) ──────┼──► evolution-session
│ L4: Architecture overhaul (user-authorized) │ (independent /evolve)
└──────────┬──────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────┐
│ Pain Detection Infrastructure │
│ pain-tracker.sh → .pain_ledger.json │
│ dirty-tracker.sh → ARCHITECTURE_MAP [STALE] │
│ overnight-plan.md → L3 work orders │
└──────────┬──────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────┐
│ Methodology Binding (non-negotiable) │
│ Every L2/L3 success MUST update │
│ retrieval-methodology routing rules │
└─────────────────────────────────────────────────┘
EN: The system can detect its own retrieval friction (repeated failed searches, stale caches) and autonomously heal — from zero-cost logging (L1) up to forging entirely new tools (L3) in dedicated evolution sessions. Every improvement is permanently bound into methodology routing rules.
CN: 系统能自主检测检索摩擦（重复失败搜索、缓存过期），并自主修复——从零成本记录（L1）到在独立进化会话中锻造全新工具（L3）。每次改进都永久绑定到方法论路由规则中。
- Delta Gate: New weapon must score strictly > current highest in same niche
- Highlander: Only one weapon per niche — old weapon physically deleted on replacement
- Forced Bootstrap: Must route through ARMORY.md to highest-tier weapon, no raw commands
- Mutation Checkpoint: Each iteration must produce real diff or declare "capped + reason"
Step 1: Full Scan / 全量扫描
2,137 community assets across 23 repos
8 parallel scanning agents
Step 2: Niche Analysis / 生态位分析
~910 new niches identified
4 red-alert competitors found
Step 3: Wave 1 Forge / 第一波锻造
8 parallel forging agents
→ 11 Skills + 1 Agent + 4 Rules + 2 Hooks
Step 4: Wave 2 Red Alert Upgrade / 第二波红色警报升级
4 parallel upgrade agents
→ 2 upgrades + 1 new Skill + 1 new Agent
Step 5: Wave 3 Supplement / 第三波补充
→ mcp-development Skill
Step 6: Verification / 二轮验证
15 dimensions × Top 10 cross-validation
→ Confirmed: no major gaps
| Repo | Purpose |
|---|---|
| claude-setup | The arsenal (this repo) / 武器库（本仓库） |
| claude-distillery | Search infrastructure: 2,137 community assets indexed / 检索基础设施 |
# Clone / 克隆
git clone git@github.com:themarshs/claude-setup.git
cd claude-setup
# The .claude/ directory contains all Skills, Agents, Rules, and Hooks.
# Claude Code will automatically detect them when you open a session
# in this directory.
#
# .claude/ 目录包含所有 Skills、Agents、Rules 和 Hooks。
# 在此目录下启动 Claude Code 会话时会自动加载。- Claude Code CLI installed
- Node.js >= 18
- Git configured with SSH
To enable community asset search via MCP:
# Clone distillery / 克隆蒸馏器
git clone git@github.com:themarshs/claude-distillery.git
cd claude-distillery && npm install
# Register in .mcp.json (already configured in this repo)
# 在 .mcp.json 中注册（本仓库已配置）MIT