Claude Code Projects評測:分身開發,一天上限200條 | Claude Code Projects Review: Parallel Threads, 200/Day Cap
By Kit 小克 | AI Tool Observer | 2026-09-18
🇹🇼 Claude Code Projects評測:分身開發,一天上限200條
Claude Code Projects 是 Anthropic 在 2026 年 9 月 17 日推出的 Beta 新功能,把原本一問一答的 Claude Code 對話,改成「你講需求,Claude 自己拆成多條執行緒平行處理」的模式。官方把這個互動比喻成「跟幕僚長交代工作」:你在 Project 裡說一次要做什麼,Claude 判斷要開新執行緒還是丟給既有執行緒,自己排程、review、最後整合結果。
Claude Code Projects 怎麼運作
核心是一個「協調者(coordinator)」加上多條「執行緒(threads)」。協調者用低算力(low effort)跑,負責拆解任務、分派工作;每條執行緒預設用 Opus 高算力跑,各自在雲端開一個獨立 session,配一個獨立 git branch 和獨立的 repo 副本,彼此不會互相覆蓋修改。連結 repo 的執行緒會自己開 PR、跑測試;連結文件的執行緒則會讀資料、產草稿。所有執行緒共享一個 Project 記憶庫(project memory)和素材庫(library),你關電腦、放著不管,Claude 照樣在雲端繼續跑,甚至可以用手機下指令調整方向。
誰能用、額度多少
- 開放對象:目前僅開放部分 Pro 與 Max 訂閱戶的雲端 session 使用,Team、Enterprise 之後才會陸續開放。
- 執行緒上限:每位訂閱戶每天最多可開 200 條新執行緒,這是唯一的硬性上限;使用者自己設的規則只是「偏好」,Claude 會盡量遵守但不是強制。
- 算力消耗:因為每條執行緒預設吃 Opus 高算力,平行開多條會比單一 session 更快把方案額度用完,這點官方文件有特別提醒。
實際使用前該注意什麼
對日常用 Claude Code 開發的人來說,這個功能解決的是「單一對話做不了太多平行任務」的痛點——以前要嘛自己開多個終端機分頭跑,要嘛乾等一個任務做完才做下一個。現在協調者幫你排程,理論上省下不少人工切換的時間。但代價很直接:Opus 高算力 × 多條執行緒,額度燒得比想像快,尤其 Max 方案的重度使用者可能幾天就摸到 200 條的上限。
目前還是 Beta,且限定部分 Pro/Max 帳號的雲端 session,一般用戶還看不到這個介面。如果你的帳號有資格,建議先拿一個小型、範圍明確的任務(例如同時修三個獨立的小 bug)試水溫,觀察額度消耗速度,再決定要不要把大型重構丟進去平行跑。
好不好用,試了才知道。
🇺🇸 Claude Code Projects Review: Parallel Threads, 200/Day Cap
Claude Code Projects is a beta feature Anthropic launched on September 17, 2026, turning Claude Code from a single back-and-forth conversation into a system where you describe what needs doing and Claude splits the work across parallel threads on its own. Anthropic compares it to briefing a chief of staff: you state the goal once inside a Project, and Claude decides whether to spin up a new thread or route the task to an existing one, then coordinates, reviews, and assembles the result.
How Claude Code Projects Actually Works
The architecture has two parts: a low-effort coordinator that scopes and delegates work, and multiple threads that do the actual work, each running Opus at high effort by default. Every repo-based thread is a full cloud session with its own branch and its own copy of the repository, so parallel threads can't overwrite each other's changes. A thread connected to a repo opens pull requests and runs tests; a thread connected to documents reads them and drafts output. All threads pull from and contribute to a shared project memory and a library of files and artifacts, and work keeps running in the cloud after you close your laptop — you can even steer it from your phone.
Who Gets Access, and the Daily Cap
- Availability: Currently in beta for select Claude Pro and Max subscribers in cloud sessions only; Team and Enterprise plans are coming later.
- Thread limit: The only hard ceiling is 200 new threads per day per subscriber. Any limit you ask Claude to follow is treated as a preference, not an enforced rule.
- Compute cost: Since every thread defaults to Opus at high effort, running several threads in parallel burns through plan usage noticeably faster than a single session — Anthropic flags this explicitly in its docs.
What to Watch Before You Rely On It
For people who already live inside Claude Code, this targets a real pain point: a single conversation couldn't handle multiple independent tasks at once, so you either juggled several terminals yourself or waited for one task to finish before starting the next. The coordinator now handles scheduling for you, which should cut down on manual context-switching. The trade-off is equally direct: Opus at high effort, multiplied across threads, eats your usage quota faster than expected — heavy Max users could hit the 200-thread daily cap within days.
It's still beta, gated to specific Pro/Max cloud accounts, so most users won't see this UI yet. If your account qualifies, start with a small, well-scoped task — like fixing three unrelated small bugs at once — to gauge quota burn before throwing a large refactor at it in parallel.
You won't know until you try it.
Sources / 資料來源
- Anthropic Launches Claude Code Projects in Beta (MarkTechPost)
- Anthropic launches Claude Code Projects, an always-on conversation (VentureBeat)
- Claude Code Projects Runs Parallel Agents on Opus by Default (Implicator.ai)
延伸閱讀 / Related Articles
- FAA空管AI評測:8.75億美元砸新創,能解決缺人危機? | FAA Air Traffic AI Review: $875M Bet vs. Controller Crisis
- AI Agent作弊評測:100個群體,27分鐘全數淪陷 | AI Agent Cheating Review: 100 Agents, 27-Minute Collapse
- Fable 5.1密碼評測:44分鐘解370年懸案,隔天被打臉 | Fable 5.1 Cipher Review: 44-Min Solve, Disputed Next Day
AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends
留言
張貼留言