一個 Claude Code 插件，把任何專案變成多代理 ideation、路線圖、QA 流水線 — 讓你不用再一直手動戳程式碼。 A Claude Code plugin that turns any project into a multi-agent ideation, roadmap, and QA pipeline — so you don't have to keep poking at the codebase by hand.
想像 Anthropic 的實習生 Claude 坐在你的終端機裡。auto-claude-ideation-suite 給這個實習生一塊寫字板和 17 位虛擬專家 — 各自負責效能、安全、程式品質、UI/UX、文件、競品分析、QA 驗證等 — 然後對他們說：去看看這個專案，告訴我下一步該做什麼。
Imagine you have an Anthropic intern called Claude sitting in your terminal. auto-claude-ideation-suite hands that intern a clipboard and a team of 17 imaginary specialists — one each for performance, security, code quality, UI/UX, documentation, competitor analysis, QA verification, and more — and tells them: go look at this project, and tell me what to do next.
輸出包含結構化的點子集、有優先級的路線圖，以及（當你做完某個功能時）一個 QA 循環，會驗證實作是否真的符合驗收條件。它會跨 session 學習 — 哪些有效、哪些失敗、哪些已完成 — 所以你不會看到重複的建議。 The output is a structured set of ideas, a prioritised roadmap, and (when you implement something) a verification loop that proves the work actually meets the acceptance criteria. It learns across sessions — what worked, what didn't, what's already done — so you don't get the same suggestion twice.
insight-extractor 代理把本 session 學到的東西存進 graph_hints.json。下次 session，所有代理都會先讀它 — 不會重複建議、會基於之前的脈絡繼續做。
An insight-extractor agent saves what was learnt this session into graph_hints.json. Next session, every agent reads it first — no duplicate suggestions, builds on prior context.
brew install claude），登入一次即可。這個插件就跑在你現有的 Claude Code 環境裡。
If you've never used Claude Code, install it from claude.ai/code (or run brew install claude) and authenticate once. The plugin runs inside your existing Claude Code setup.
npm install。直接 clone 就好。
The compiled MCP server is committed to the repo, so you don't need to run npm install. Just clone.
/auto-claude-ideation-suite、/ideation、/roadmap、/qa。
On next launch, Claude Code auto-discovers the plugin's commands, agents, skill, and the MCP server. You should see /auto-claude-ideation-suite, /ideation, /roadmap, /qa in the slash command menu.
/，找這四個命令。看到了就 OK。
Open Claude Code in any project, type /, and look for the four commands. If you see them — you're ready.
ideation、roadmap、all、qa <id>、plan <id>，或單一代理如 performance。不帶參數的話，跑全部 6 個 ideation 代理。
Pass an arg to scope: ideation, roadmap, all, qa <id>, plan <id>, or a single agent like performance. Without an arg, runs all 6 ideation agents.
ideas_merged.json。視專案大小，1–3 分鐘。
Performance, security, code-quality, code-improvements, documentation, ui-ux. Each one writes its own ideas JSON, then the plugin merges them into ideas_merged.json. Takes 1–3 minutes depending on project size.
/ideation。
Ideation answers "what could we do?" — Roadmap answers "what should we do, and in what order?" — QA answers "did we do it correctly?" Run them in that order on a fresh project; on subsequent sessions, just /ideation when you need fresh ideas.
每次執行會在專案根目錄的 .aperant/ 下建一個時間戳資料夾。代理輸出的所有 JSON 都在裡面，可以 grep、diff、跟程式碼一起 version control。
Each run creates a timestamped folder under .aperant/ in the project root. Everything the agents produce ends up there as JSON, so you can grep, diff, and version-control the outputs alongside your code.
your-project/ ├── .aperant/ │ └── 2026-04-26T10-30-00/ │ ├── project_index.json — 掃描的專案內容what we scanned │ ├── performance_ideas.json — 各代理點子per-agent ideas │ ├── security_ideas.json │ ├── code_quality_ideas.json │ ├── code_improvements_ideas.json │ ├── documentation_ideas.json │ ├── ui_ux_ideas.json │ ├── ideas_merged.json — 統一輸出unified output │ ├── roadmap_discovery.json — 若跑了 /roadmapif /roadmap was run │ ├── competitor_analysis.json │ ├── roadmap.json — MoSCoW 功能清單MoSCoW features │ └── task-state.json — 計畫步驟追蹤plan step tracking └── graph_hints.json — 跨 session 記憶cross-session memory
插件包含 17 個子代理（每個是含 YAML frontmatter 的 Markdown 提示檔）、10 個 MCP 工具（一個小型 TypeScript 伺服器）、1 個 skill（編排協定）、4 個斜線命令（入口）。當你輸入斜線命令，Claude Code 讀取它的指令，然後驅動 skill；skill 呼叫 MCP 伺服器去 scan、驗證、合併，並呼叫所需的子代理。 The plugin ships 17 sub-agents (each a Markdown prompt file with YAML frontmatter), 10 MCP tools (a small TypeScript server), 1 skill (the orchestration protocol), and 4 slash commands (the entry points). When you type a slash command, Claude Code reads its instructions, then drives the skill, which calls the MCP server to scan, validate, and merge — and spawns whichever sub-agents are needed for the job.
想看每一層、每個工具的完整架構圖，去看第三篇 auto-claude-ideation-suite-fix.html — 那是技術審計版。這頁是友好導覽版。
If you want the deep architecture diagram with every layer and tool, open the companion auto-claude-ideation-suite-fix.html — that's the technical audit. This page is the friendly intro.
node --version 確認；太舊的話用 nvm 或 brew 升級。
The MCP server needs Node 18+. Run node --version to confirm. If older, upgrade via nvm or brew.
agents/ 放一個含 YAML frontmatter (name/description/tools) 的新 .md 檔，更新 plugin.json，發 PR。CI 會在每次 push 檢查結構。
Drop a new .md file in agents/ with the YAML frontmatter (name/description/tools), update plugin.json, open a PR. CI checks structure on every push.
/ideation 跑在 1–2k LOC 的 codebase 大約一分鐘，先感受輸出長什麼樣，再放到正式專案上。
Pick a personal side project to start. /ideation on a 1–2k LOC codebase finishes in about a minute and gives you a feel for what the output looks like before you let it loose on production.