# 策展 · X (Twitter) 🔥🔥🔥

> 📖 本站完整內容索引（documentation index）：[llms.txt](/llms.txt)

> 作者：MIKE (@mikenevermiss) · 平台：X (Twitter) · 日期：2026-07-05

> 原始來源：https://x.com/mikenevermiss/status/2073278341377912944

## 中文摘要

Claude Fable 5 屬於 Mythos-class 等級，專為自主任務設計。

Claude Fable 5 是 Anthropic 目前對外公開最強大的模型，屬於全新的「Mythos-class」等級，專為處理過去無法輕易委派、需耗時數週的複雜自主工作而設計。該模型於 2026 年 6 月 9 日發布，雖然曾因出口管制報告短暫停用，但已於 7 月 1 日全球恢復存取。

**核心能力與應用場景**
Fable 5 的優勢在於處理長週期、多步驟的任務，而非單純的快速問答。
- **工程遷移**：Stripe 曾利用 Fable 5 在一天內完成 5,000 萬行 Ruby 程式碼庫的遷移，此類任務若由人工執行需耗時兩個月以上。
- **深度分析**：在 Hebbia 的金融基準測試中，Fable 5 展現了頂尖的資深層級推理能力，並能精準處理複雜的圖表、科學數據與視覺資訊。
- **記憶與脈絡**：具備 100 萬 token 的 context window，並能在任務執行期間自我記錄筆記，以提升長執行過程中的輸出品質。

**安全系統與分類器機制**
Fable 5 內建獨立的 AI 分類器，用以監控請求並過濾高風險內容。
- **觸發機制**：若請求涉及網路安全（漏洞挖掘）、生物與化學（實驗方法、基因內容）或模型蒸餾，系統會自動將請求轉由 Claude Opus 4.8 處理。
- **應對策略**：若被轉向，使用者可透過增加背景脈絡（說明為何需要此資訊）或明確定義良性用途來重新嘗試。
- **注意事項**：避免在 prompt 中要求模型「展示思考過程」或「敘述推理步驟」，這類指令容易觸發推理提取分類器，導致系統自動降級至 Opus。

**存取方式與 effort levels**
使用者可透過 `claude.ai`、`Claude Code`、`Claude Cowork` 或 API（模型代號 `claude-fable-5`）存取。
- **effort levels**：這是控制模型智慧、速度與成本的關鍵參數，分為 `low`、`medium`、`high` 與 `xhigh`。
    - `low` / `medium`：適用於日常摘要或草稿，速度較快。
    - `high`：大多數實際工作的預設值，模型會進行規劃、執行並自我驗證。
    - `xhigh`：適用於最困難的任務，模型會進行更深度的審慎思考與自我驗證。
- **操作技巧**：在 `claude.ai` 中，可透過自然語言指令（如「take your time, check your reasoning carefully」）來模擬高 effort 模式。

**實用 Prompt 範本**
以下為針對不同場景設計的 prompt，建議直接複製使用：

1. **長文件分析**：
   ```
   I'm preparing for a client presentation on [topic]. the attached document is the primary source. my goal is to identify the three most consequential findings and any points where the evidence is weak or contradictory. do not summarize the whole document. go section by section, flag what matters most, and tell me when something needs more verification before I cite it. stop and check with me before making any claims you're not confident about.
   ```

2. **多來源研究簡報**：
   ```
   I'm building a research brief on [topic] for [audience]. I've attached [X sources]. my goal is a structured brief that identifies where the sources agree, where they conflict, and what questions remain unanswered. do not write a summary of each source separately. synthesize across all of them. flag any gaps that would materially affect the conclusion. when you're done, tell me what additional sources would strengthen the brief.
   ```

3. **長篇內容撰寫**：
   ```
   I'm writing a [type of content] for [audience]. my goal is a complete first draft, not an outline. the piece should [describe purpose]. here is an example of the voice and format I want: [paste 2-3 paragraphs of your own writing]. match that voice exactly. if you're uncertain whether a section serves the purpose, flag it rather than cut it. deliver the full draft, then list any sections you were uncertain about at the end.
   ```

4. **專案規劃**：
   ```
   I'm planning [project] with a team of [X] people and a deadline of [date]. my goal is a realistic phased plan, not an ideal one. identify where the schedule is tight, what dependencies could cause delays, and what I should cut first if time runs short. do not give me a generic project template. use the constraints I've given you. at the end, flag the two or three decisions that will most affect whether this succeeds.
   ```

5. **數據分析**：
   ```
   I'm analyzing [type of data] to [business goal]. I've attached [file]. my goal is to identify patterns that would not be obvious to someone looking at the raw numbers. start by describing what the data actually contains and any quality issues you notice. then give me the two or three findings most relevant to [goal]. flag anything I should not conclude from this data alone. stop and check with me if you find something that contradicts what I told you to expect.
   ```

6. **自主任務（Claude Code）**：
   ```
   I'm working on [the larger task] for [who it's for]. they need [what the output enables]. proceed end to end without checking in unless you hit a destructive or irreversible action, a real scope change, or something only I can provide. for those cases, pause and tell me exactly what you need. before reporting any progress, verify each claim against actual results. if something failed, say so plainly. do not end your turn with a plan or a list of next steps you haven't done yet. finish the work.
   ```

7. **回饋與編輯**：
   ```
   I have a draft of [type of content]. I want specific feedback, not a rewrite. for each issue you find, quote the sentence or paragraph, identify what's wrong, and suggest the precise fix. do not improve things that are working. do not give me general notes like "this section could be stronger." if a section is weak, tell me exactly why and exactly what to change. flag anything where you're uncertain whether it's a problem or a stylistic choice.
   ```

8. **決策分析**：
   ```
   I'm deciding [decision]. the relevant context is [situation]. I'm considering [option A] versus [option B]. I'm asking for analysis, not agreement. identify the strongest argument for each option, the biggest risk I'm probably underweighting, and what additional information would most change the decision. do not tell me what you think I want to hear. if the right answer is "you don't have enough information yet," say that. end with one clear recommendation and the one thing I should verify before acting on it.
   ```

**常見錯誤與建議**
- **勿大材小用**：不要將 Fable 5 用於簡單的郵件摘要或快速問答，這類任務使用 Sonnet 5 或 Haiku 4.5 更為高效且經濟。
- **給予足夠時間**：在高 effort 設定下，模型需要時間進行規劃與自我驗證，請勿在短時間內中斷執行。
- **明確終止條件**：若未在 prompt 中設定明確的停止條件，模型可能會進行過度延伸的處理，務必告知模型「何謂完成」。

## 標籤

Agent, 新產品, 功能更新, Claude, Anthropic, Claude