Claude Code的實用指令
初始架構
/init: 分析當前專案目錄,自動偵測技術棧並建立基礎設定。
掃描整個 codebase,在專案根目錄生成 CLAUDE.md(如果已存在則更新),內容包含:
- 技術架構摘要(Next.js、Prisma、TypeScript…)
- 常用指令(
npm run build:frontend、npm run build:backend…) - 專案慣例(DB 讀寫分離規則、資料夾結構…)
- Claude 每次開新 session 會自動讀取它,不需要重新說明背景
如何安裝Claude Code CLI
透過 PowerShell
irm https://claude.ai/install.ps1 | iex
安裝完成後關閉PowerShell,並重新開啟,執行 claude 即可啟動
若出現 claude is not recognized 錯誤,手動加入 PATH
[Environment]::SetEnvironmentVariable("PATH", "$env:PATH;$env:USERPROFILE\.local\bin", [EnvironmentVariableTarget]::User)
以上,即完成安裝。
程式碼審查
/review: 對當前檔案或 Git 變更進行代碼品質審查。
/undo (或 /rewind): 撤銷上一步 AI 所做的代碼修改。
/iterate
technical-writing-improver.SKILL.md
# SKILL.md using Iterative Refinement pattern
# Example technical-writing-improver skill
You improve technical writing through iterative refinement.
Your workflow has explicit quality gates at each iteration.
## Iteration 1: Structural review
- Read the draft and check overall structure
- Verify introduction states the problem clearly
- Check sections follow logical flow
- Ensure conclusion summarizes key points
- Score structure 1-10 and explain reasoning
## Iteration 2: Clarity review
- For each paragraph check readability score
- Identify sentences with Flesch-Kincaid above grade 12
- Suggest simpler phrasings without losing accuracy
- Score clarity 1-10 and list improvements
## Iteration 3: Technical accuracy
- Verify all code examples compile and run
- Check API references against current documentation
- Validate command flags and options exist
- Cross-reference version numbers
- Flag any factual errors found
## Iteration 4: Examples and context
- Every technical concept should have example
- Examples should use realistic data not foo bar
- Add common pitfalls section if missing
- Check external links are not broken
## Validation script for each iteration
Run `scripts/validate-markdown.py content.md`
- Reports issues and blocks progression if severity high
- Provides numeric score for tracking
## Stop iterating when
- All iterations reach score 8+
- No validation script failures
- User approves quality gate
Do not iterate indefinitely.
Maximum 5 total iterations.
使用方式:/iterate improve docs/content.md using technical-writing-improver
執行流程:
- Claude載入SKILL.md(Progressive Disclosure,只在相關時載入)。
- Iteration 1:結構審查,得分9/10,輸出改版content.md。
- 執行
validate-markdown.py,若通過進Iteration 2;否則重試。 - 依序進行至所有階段≥8分,或達5次上限停止。
- 總Token成本預測性高(每階段固定),品質優於單次prompt。
Token節省
/plan: 進入「計畫模式」,只進行邏輯規劃而不實際修改代碼(節省 Token)。
/exit_plan: 部分版本支援直接輸入指令退出計畫模式並返回一般模式。
/rewind: 如果你發現計畫大錯特錯,可以用此指令回到計畫前的狀態重新開始。
/clear: 完成任務後,可以透過這個指令清除之前的Context。
/compact: 壓縮當前對話歷史,釋放上下文空間(Context Window)。
/effort low:
/effort high:
/btw:討論中,如果有想要詢問其他議題,可以透過這個指令避免這個討論被加入session memory。
GitHub – forrestchang/andrej-karpathy-skills
A single CLAUDE.md file to improve Claude Code behavior, derived from Andrej Karpathy’s observations on LLM coding pitfalls. · GitHub
