Back to Details

cogworks

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Cogworks

Cogworks

Overview

概述

You are the single product entry point for turning source material into a trustworthy generated agent skill.
Optimize for:
  • skill quality
  • source trustworthiness
  • concise user-facing flow
  • minimal context pollution
The generated skill is the product artifact. Runtime machinery exists only to improve that artifact.
你是将源材料转化为可信的生成式Agent技能的唯一产品入口。
优化方向：
  • 技能质量
  • 源可信度
  • 简洁的面向用户流程
  • 最小化上下文污染
生成的技能是产品产物，运行时组件仅用于优化该产物。

When to Use

适用场景

Use this skill only when the user explicitly invokes 
cogworks
and wants skill generation as the outcome.
If the request is analysis-only, manual skill-writing help, or does not clearly ask for generation, clarify before creating files.
If the user asks what cogworks is, how to use it, or what support boundaries exist, read README.md.
仅当用户显式调用
cogworks
且期望生成技能作为输出时，才使用该技能。
如果请求仅涉及分析、手动编写技能的帮助，或未明确要求生成技能，请先澄清需求再创建文件。
如果用户询问cogworks是什么、如何使用，或适用边界是什么，请查阅README.md。

Quick Decision Cheatsheet

快速决策速查表

  • explicit 
    cogworks
    invocation means generation intent is already established
  • verify 
    cogworks-encode
    and 
    cogworks-learn
    before running
  • trust classification happens before synthesis, never after
  • unsupported surfaces fail closed rather than degrading silently
  • fail closed when trust, provenance, contradiction handling, or validation is insufficient.
  • only the generated skill is a user-facing product artifact
  • deterministic validation is a hard gate before final output
  • runtime details such as execution surface, run root, or sub-agent metadata do not belong in generated skill frontmatter or metadata
  • 显式调用
    cogworks
    即表明已明确生成意图
  • 运行前先验证
    cogworks-encode
    和
    cogworks-learn
    可用
  • 可信度分类必须在合成之前完成，不可后置
  • 不支持的场景主动报错终止，而非静默降级
  • 当可信度、来源追溯、矛盾处理或验证不足时，主动报错终止
  • 仅生成的技能是面向用户的产品产物
  • 确定性验证是最终输出前的硬性门槛
  • 运行时详情（如执行环境、运行根目录、子Agent元数据）不得出现在生成技能的前置信息或元数据中

Invocation

调用方式

Use 
cogworks
to:
  • verify both dependency skills are present and readable
  • resolve topic, sources, destination, and metadata defaults
  • build 
    dispatch-manifest.json
    from 
    role-profiles.json
    as the canonical source for 
    binding_ref
    , 
    model_policy
    , 
    preferred_dispatch_mode
    , and the canonical 
    tool_scope
    string
  • after the specialist dispatch modes are known, write 
    dispatch-manifest.json
    with 
    python3 scripts/render-dispatch-manifest.py --surface <surface> --output {run_root}/dispatch-manifest.json ...
    and provide per-profile 
    --actual-mode profile_id=mode
    overrides as needed
  • classify trust before synthesis using 
    cogworks-encode
  • run the fixed internal build through packaging and deterministic validation
  • apply 
    cogworks-learn
    packaging rules to the final skill
  • keep user-facing narration to one short progress line per stage
Do not invoke this skill for general documentation Q&A or manual skill-writing advice unless the user explicitly wants generation.
For the stable operator checklist, failure conditions, and stage contract, use reference.md.
When runtime adapters expose overlapping metadata, the canonical fields from role-profiles.json win over generated adapter files.
使用
cogworks
完成以下操作：
  • 验证两个依赖技能都存在且可读
  • 解析主题、来源、存储路径和元数据默认值
  • 基于
    role-profiles.json
    构建
    dispatch-manifest.json
    ，作为
    binding_ref
    、
    model_policy
    、
    preferred_dispatch_mode
    以及标准
    tool_scope
    字符串的权威来源
  • 明确专业角色调度模式后，通过
    python3 scripts/render-dispatch-manifest.py --surface <surface> --output {run_root}/dispatch-manifest.json ...
    写入
    dispatch-manifest.json
    ，按需提供
    --actual-mode profile_id=mode
    参数按角色覆盖模式
  • 合成前通过
    cogworks-encode
    完成可信度分类
  • 执行固定内部构建流程，完成打包和确定性验证
  • 对最终技能应用
    cogworks-learn
    打包规则
  • 面向用户的进度说明每个阶段仅保留一行简短提示
除非用户明确要求生成技能，否则不要为常规文档问答或手动编写技能建议调用该技能。
如需查看稳定操作清单、故障场景和阶段约定，请查阅reference.md。
当运行时适配器存在重复的元数据时，role-profiles.json中的权威字段优先级高于生成的适配器文件。

Compatibility

兼容性

Claude Code enforces the manual-only posture for this skill via 
disable-model-invocation: true
.
Codex enforces the same posture via agents/openai.yaml, with implicit invocation disabled.
Other runtimes may ignore these platform-specific controls. Keep treating explicit user invocation as the policy boundary for any run that can create files or directories.
Claude Code通过
disable-model-invocation: true
强制该技能仅支持手动调用。
Codex通过agents/openai.yaml强制执行相同规则，隐式调用已被禁用。
其他运行时可能忽略这些平台专属控制，对于任何可创建文件或目录的运行，始终将用户显式调用作为策略边界。

Supporting Docs

配套文档

  • README.md: user-facing product overview and support boundaries
  • reference.md: stable product contract and operator checklist
  • metadata.json: repo-local release metadata for this skill
  • agents/openai.yaml: Codex-specific invocation policy
  • agentic-runtime.md: maintainer-only runtime contract
  • claude-adapter.md: Claude-specific maintainer guidance
  • copilot-adapter.md: Copilot-specific maintainer guidance
  • role-profiles.json: canonical specialist role bindings
The frontmatter 
metadata
block is a repo-local convention. Other platforms may ignore it; canonical package metadata for tooling lives in metadata.json.
  • README.md：面向用户的产品概述和支持边界
  • reference.md：稳定产品约定和操作清单
  • metadata.json：该技能的仓库本地发布元数据
  • agents/openai.yaml：Codex专属调用策略
  • agentic-runtime.md：仅维护者可见的运行时约定
  • claude-adapter.md：Claude专属维护指南
  • copilot-adapter.md：Copilot专属维护指南
  • role-profiles.json：权威专业角色绑定配置
前置信息中的
metadata
块是仓库本地约定，其他平台可能会忽略该内容；工具使用的标准包元数据存储在metadata.json中。