Agent审查
- 作者仓库星标 3,406
- 作者更新于 实时读取
- 作者仓库 claude-octopus
- 领域
- 写作
- 兼容 Agent
-
- Claude Code
- Cursor
- Cline
- Codex
- Windsurf
- Gemini CLI
- +20
- 信任分
- 88 / 100 · 社区维护
- 作者 / 版本 / 许可
- @nyldn · 未声明 license
- Token 消耗评级
- 低消耗
- 接入复杂程度
- 需简单配置
- 是否需要外部 API Key
- 不需要
- 兼容的系统
- macOS · Linux · Windows
- 底层运行要求
- Node.js
- 文件与系统权限
-
- 只读
- 允许写入 / 修改
- Shell 执行
- 网络行为
- 仅限本地
- 安装命令数
- 26 条
档案由构建时根据 SKILL.md 与安装命令自动衍生，可能与作者实际意图存在差异。
需要注意： 未限定 allowed-tools，默认拥有全部工具权限。
---
name: skill-knowledge-work
description: Switch to Knowledge Work mode for research and writing — use when task is non-code focused All s…
category: 写作
runtime: Node.js
---
# skill-knowledge-work 输出预览
## PART A: 任务判断
- 适用问题：文章、文案、发言稿、润色或结构化表达。
- 输入要求：目标材料、限制条件、期望输出和验收方式。
- 证据边界：围绕“Context Auto-Detection (v7.8+) / When to Use This Override / Override Commands”读取原文规则，不把推断写成作者承诺。
## PART B: 执行结果
- **01** 任务判断：确认你的需求是否属于文章、文案、发言稿、润色或结构化表达，并标出输入、限制和预期结果。
- **02** 执行计划：优先按“Context Auto-Detection (v7.8+) / When to Use This Override / Override Commands”拆成步骤，说明每一步会读取什么、修改什么、产出什么。
- **03** 交付结果：给出可复制的命令、文件改动、检查清单或内容草稿，并说明如何继续迭代。
- **04** 风险边界：结合 读取文件、写入/修改文件、执行终端命令、主要在本地完成、通常不需要额外 API Key 给出执行前确认项。
## Running Rules
- 读取文件、写入/修改文件、执行终端命令；主要在本地完成；通常不需要额外 API Key。
- 先小样例验证，再放大到真实任务。
- 交付时同时给结果、检查口径和下一步迭代建议。原文出现了 `/octo` 这类斜杠命令；如果你的 Agent 支持命令触发，优先用命令开场，再补充目标和边界。
告诉 Agent 目标文件或材料、期望结果、不可改范围、是否允许联网或执行命令。本 Skill 的权限画像是：读取文件、写入/修改文件、执行终端命令。
先用一个小任务确认它会围绕“Context Auto-Detection (v7.8+) / When to Use This Override / Override Commands”工作；涉及文件或命令时，先看 diff、日志、预览或测试结果。
检查最终产物是否包含明确结果、必要证据和下一步动作；如果输出泛泛而谈，就补充输入、边界和验收标准后重跑。
---
name: skill-knowledge-work
description: Switch to Knowledge Work mode for research and writing — use when task is non-code focused All s…
category: 写作
source: nyldn/claude-octopus
---
# skill-knowledge-work
## 什么时候使用
- 把写作方向的常用动作沉淀成 Agent 可调用的技能 适合处理文章、文案、润色、翻译、总结和结构化表达，核心价值是把输入、判断、执行、验证和交付边界固定下来，避免 Agent 泛泛回答。 把任务拆成可执行、可检查、可继续迭代的步骤；通常…
- 面向文章、文案、发言稿、润色或结构化表达，优先处理能明确输入、步骤和验收标准的工作。
## 需要提供什么
- 目标材料、目录范围、期望结果和不可改动内容。
- 是否允许联网、执行命令、读写文件或调用外部服务。
## 执行规则
- 围绕「Context Auto-Detection (v7.8+) / When to Use This Override / Override Commands」组织步骤，不把推断写成作者事实。
- 读取文件、写入/修改文件、执行终端命令；主要在本地完成；通常不需要额外 API Key。
- 先跑小样例，确认结果可检查后再扩大任务范围。
## 输出要求
- 给出最终产物、关键证据、验证方式和下一步动作。
- 信息不足时标记 unknown，不编造命令、平台或依赖。作者原文负责流程事实；仓库文件负责来源和命令；流狐只补充适用场景、限制和质量判断。
skill "skill-knowledge-work" {
输入层 -> 用户目标 + 目标文件 + 禁止范围 + 验收标准
上下文层 -> Context Auto-Detection (v7.8+) / When to Use This Override / Override Commands
规则层 -> SKILL.md 触发条件 / 执行顺序 / 输出格式
运行层 -> Node.js | 读取文件、写入/修改文件、执行终端命令 | 主要在本地完成
安全层 -> 通常不需要额外 API Key + 小任务验证 + diff / 日志复核
输出层 -> 可复制结果 + 检查清单 + 下一步迭代
}Host: Codex CLI — This skill was designed for Claude Code and adapted for Codex. Cross-reference commands use installed skill names in Codex rather than
/octo:*slash commands. Use the active Codex shell and subagent tools. Do not claim a provider, model, or host subagent is available until the current session exposes it. For host tool equivalents, seeskills/blocks/codex-host-adapter.md.
Knowledge Work Mode - Context Override Skill
Context Auto-Detection (v7.8+)
Claude Octopus now auto-detects work context! The system analyzes your prompt and project to determine whether you're in a Dev Context (code-focused) or Knowledge Context (research/strategy-focused).
You typically don't need this skill - context is detected automatically when you use:
octo research X- Auto-detects dev vs knowledge researchocto build X- Auto-detects code vs document buildingocto review X- Auto-detects code vs document review
When to Use This Override
Use ONLY when auto-detection is wrong:
- Auto-detection chose Dev but you want Knowledge behavior
- Auto-detection chose Knowledge but you want Dev behavior
- You want to force a specific context for the entire session
Override Commands
Force Knowledge Context
/octo:km onAll subsequent workflows will use Knowledge Context until reset.
Force Dev Context
/octo:km offAll subsequent workflows will use Dev Context until reset.
Return to Auto-Detection
/octo:km autoContext detection returns to automatic mode.
Check Current Status
/octo:kmShows current mode (auto, knowledge, or dev).
How Auto-Detection Works
When you use any octo workflow, context is detected by analyzing:
Prompt Content (strongest signal):
- Knowledge indicators: "market", "ROI", "stakeholders", "strategy", "personas", "presentation", "report", "PRD"
- Dev indicators: "API", "endpoint", "database", "implementation", "code", "function", "deploy"
Project Type (secondary signal):
- Has
package.json,Cargo.toml,go.mod→ Dev Context - Mostly
.md,.docx,.pdffiles → Knowledge Context
- Has
Explicit Override (if set via
/octo:km):- Overrides all auto-detection until reset to "auto"
Visual Indicator Shows Context
When workflows run, you'll see the detected context in the banner:
Dev Context:
🐙 **CLAUDE OCTOPUS ACTIVATED** - Multi-provider research mode
🔍 [Dev] Discover Phase: Technical research on caching patternsKnowledge Context:
🐙 **CLAUDE OCTOPUS ACTIVATED** - Multi-provider research mode
🔍 [Knowledge] Discover Phase: Market analysis for APAC expansionExamples of Auto-Detection in Action
Example 1: Technical Research (Auto → Dev)
User: "octo research caching strategies for our Node.js API"
Claude: (auto-detects Dev Context from "Node.js API")
🐙 **CLAUDE OCTOPUS ACTIVATED** - Multi-provider research mode
🔍 [Dev] Discover Phase: Technical research on caching strategies
[Researches with technical/implementation focus]Example 2: Market Research (Auto → Knowledge)
User: "octo research market opportunities in healthcare AI"
Claude: (auto-detects Knowledge Context from "market opportunities")
🐙 **CLAUDE OCTOPUS ACTIVATED** - Multi-provider research mode
🔍 [Knowledge] Discover Phase: Strategic research on healthcare AI market
[Researches with business/strategic focus]Example 3: Override When Auto Gets It Wrong
User: "octo research React patterns"
Claude: (auto-detects Dev) 🔍 [Dev] Discover Phase...
User: "Actually, this is for a presentation. Force knowledge mode."
Claude: "Setting context override to Knowledge Mode."
/octo:km onUser: "octo research React patterns"
Claude: (uses override) 🔍 [Knowledge] Discover Phase... (focuses on trends, adoption, strategic implications)
Example 4: Check Current Status
User: "What context mode am I in?"
Claude:
Current mode: Auto-detection (no override set)
Last detected context: Dev (based on project having package.json)
To override: /octo:km on (force Knowledge) or /octo:km off (force Dev)
To return to auto: /octo:km autoWhat Changes Per Context
Dev Context 🔧
| Workflow | Focus |
|---|---|
octo research X |
Technical implementation, libraries, code patterns |
octo build X |
Code generation, architecture, tests |
octo review X |
Code quality, security, performance |
| Agents | codex, backend-architect, code-reviewer, security-auditor |
Knowledge Context 🎓
| Workflow | Focus |
|---|---|
octo research X |
Market analysis, competitive research, literature synthesis |
octo build X |
PRDs, strategy docs, presentations, reports |
octo review X |
Document quality, argument strength, completeness |
| Agents | strategy-analyst, ux-researcher, exec-communicator, product-writer |
Document Delivery 📄
After running knowledge workflows, export to professional formats:
- DOCX - Word documents for reports, business cases
- PPTX - PowerPoint presentations for stakeholder decks
- XLSX - Excel spreadsheets for data analysis
Just say: "Export this to Word" or "Create a PowerPoint presentation"
Override Command Reference
| Command | Description |
|---|---|
/octo:km |
Show current status (auto, on, or off) |
/octo:km on |
Force Knowledge Context for all workflows |
/octo:km off |
Force Dev Context for all workflows |
/octo:km auto |
Return to auto-detection (default) |
When NOT to Use Override
Don't override if:
- Auto-detection is working correctly
- You're doing mixed work (let each prompt be detected individually)
- You just want to see what context was detected (check the banner)
Override is for:
- Forcing a specific context for an entire session
- Correcting persistent misdetection
- Specific use cases where you know better than auto-detect
Cross-Task Learnings
At the end of significant work sessions, extract learnings:
- What task type was this? (debugging, implementation, research, review)
- What approach worked? What failed?
- What would you do differently next time?
Store learnings in .claude-octopus/learnings/<date>-<summary>.json:
{
"date": "2026-03-21",
"task_type": "debugging",
"approach": "Traced the error from the test failure back to the API handler",
"outcome": "success",
"lesson": "Always check middleware ordering before investigating handler logic"
}Each learning file captures: task_type, approach, outcome, and lesson.
Session Start: Relevance Matching
At session start, check for relevant learnings:
- Read
.claude-octopus/learnings/directory - Match by task type and file patterns relevant to the current work
- Inject top 3 most relevant learnings as context (within ~5% token budget)
- Prefer recent learnings; deprioritize those older than 30 days
Budget Controls
- Maximum 5 learnings extracted per session (prevents runaway writes)
- Maximum 50 learning files retained (oldest pruned automatically)
- Relevance injection capped at 3 learnings and ~5% of available token budget
- Learning files are lightweight JSON (~200 bytes each)
Related Skills
/octo:discover- Research workflow (auto-detects context)/octo:develop- Build workflow (auto-detects context)/octo:deliver- Review workflow (auto-detects context)/octo:docs- Document export (works in both contexts)
先判断是否适合
作者设计意图
作者的方法与取舍
边界和复核