Token导航 LogoToken导航

Agent Skill · AI 工具

composable-step-design(可组合的步骤设计)

用于辅助界面设计、视觉规范、排版、配色、布局和交互体验优化。它适合让 Agent 根据产品场景整理页面结构、生成 UI 方案、检查视觉一致性或改进组件层级。使用时需要结合现有品牌、设计系统和用户任务,不应只堆装饰元素;涉及真实页面改动时,应通过截图或浏览器预览检查文本溢出、对齐和响应式表现。

composable-step-design 是AI 工具类 Agent Skill,可用宿主为 Codex、Claude、Cursor、Gemini CLI,通过GitHub 仓库获取,权限为只读;许可证为 MIT,第三方扫描已通过,按安装量在AI 工具类的 2,536 个 Skill 里排第 2,213。

总安装 282周安装 12GitHub Stars 61下载 99核验 2026-05-01

安装说明

本站只整理中文说明,不托管安装包,也不代用户安装。
GitHub
许可证
MIT

安装方式

通过对话安装

复制提示词发给支持本地命令或 Skills 的 AI 助手,先确认命令和权限,再让它执行。

请帮我安装这个 Agent Skill:composable-step-design(可组合的步骤设计)
来源仓库:https://github.com/melodic-software/claude-code-plugins
仓库路径:skills/composable-step-design
安装命令:
npx skills add https://github.com/melodic-software/claude-code-plugins --skill composable-step-design
安装前请先检查当前环境是否支持对应 CLI,并向我确认将要执行的命令、安装目录、联网范围和文件读写权限;确认后再执行。

命令行安装

复制命令到本机终端执行。不同来源提供的安装方式可能略有差异;本站展示可直接复制的安装命令,安装前请核对来源页面。

skills.shnpx skills
npx skills add https://github.com/melodic-software/claude-code-plugins --skill composable-step-design

这个 Skill 是做什么的

AI 工具类 Skill 用来扩展助手本身的能力,例如接模型、接检索、接外部服务。

composable-step-design 设计 AI 开发者工作流的四个可组合步骤:计划、构建、审查、修复。适用于创建新工作流步骤原语或定义步骤输入/输出契约的场景。支持步骤隔离模式和组合策略设计,形成稳定的 ADW 工作流变体。完整步骤编排需配合 Claude Agent SDK 使用,本技能提供基础框架。

使用前应熟悉 @composable-steps.md 内存文件和 ADW 框架模式。

来源仓库内路径:skills/composable-step-design

上手步骤

4 步
  1. 1

    确认自己的宿主工具

    公开资料标注支持 Codex、Claude、Cursor、Gemini CLI;先确认自己用的是其中之一,再按对应工具的 Skills 目录放置。

  2. 2

    拿到安装命令

    可以直接复制的命令是 npx skills add https://github.com/melodic-software/claude-code-plugins --skill c…;命令里的占位符换成自己的路径或密钥再执行。

  3. 3

    安装后重启宿主

    命令执行完要完全退出宿主工具再打开,让新装的 Skill 被重新扫描到,否则列表里可能看不到。

  4. 4

    按最小权限试用

    它标注为只读,先用只读任务试一次,确认行为符合预期再扩大使用范围。

常见问题

6 条

composable-step-design 怎么安装?

在终端执行 npx skills add https://github.com/melodic-software/claude-code-plugins --skill composable-…。装完完全退出宿主工具再打开,让它重新扫描本地 Skill。

composable-step-design 支持哪些 AI 工具?

公开资料标注支持 Codex、Claude、Cursor、Gemini CLI。其他工具只要能读取本地 Skill 目录,一般也能手动放进去,但要自己确认触发方式是否一致。

composable-step-design 需要什么权限?

权限级别是只读。该 Skill 主要提供规则、说明或参考内容,本身偏只读;真正读写文件、联网或执行命令仍取决于宿主 Agent 的任务。

composable-step-design 是免费的吗?

来源标注的许可证是 MIT。本站不代收任何费用,也不托管安装包;能否商用、要不要署名,以许可证原文为准。

composable-step-design 用的人多吗?

公开数据是累计安装 282 次,近一周安装 12 次,下载 99 次,GitHub 61 星。这些数字来自第三方目录统计,只能当作热度参考,不代表质量或安全性。

composable-step-design 安全吗?

第三方扫描结果为通过,但这只覆盖已知的静态检查项,不能替代自己读一遍源码和安装命令。

值得选的地方和要先想清楚的地方

值得选的地方

  • 同时适配 4 种工具,换客户端不用重新找替代品。
  • 3 个目录同时收录,说明这个 Skill 流传比较广。
  • 许可证明确为 MIT,商用和分发边界好判断。
  • 第三方扫描已通过,常见的高风险行为没有命中。

先想清楚的地方

  • 第三方目录整理,版本更新和行为细节以来源仓库为准。

适不适合你

更适合

  • 已经在用 Codex,想直接扩展它能力的人。
  • 需要处理AI 工具类任务、又不想每次都手写提示词的人。
  • 需要许可证清楚、并且有第三方扫描结果的团队项目。

不太适合

  • 需要厂商级技术支持和 SLA 的生产系统。

AI 工具类 Skill 对比

6 个同类
Skill总安装GitHub 星权限许可证扫描
executing-plans

执行计划 · Codex、Claude

1,515,336173,110只读MIT通过
verification-before-completion

完成前的验证 · Codex、Claude

1,319,472173,121权限未标注MIT通过
subagent-driven-development

子 Agent 驱动开发 · Codex、Claude

1,312,416173,138只读MIT通过
copy-editing

复制编辑 · Codex、Claude

1,209,57625,632只读MIT通过
paywall-upgrade-cro

付费专区升级 cro · Codex、Claude

1,030,17625,617只读MIT通过
composable-step-design

可组合的步骤设计 · 本页

28261只读MIT通过

按安装量取同分类前几名;数字来自第三方目录统计,只作热度参考,不代表质量结论。

SKILL.md 原文

Composable Step Design

Design the four composable workflow steps (Plan/Build/Review/Fix) that form the backbone of AI Developer Workflows.

When to Use

  • Creating new workflow step primitives
  • Defining step input/output contracts
  • Implementing step isolation patterns
  • Designing step composition strategies
  • Building ADW workflow variations

Prerequisites

  • Understanding of @composable-steps.md memory file
  • Familiarity with @adw-framework.md patterns
  • Knowledge of deterministic vs non-deterministic execution

SDK Requirement

Implementation Note: Full step orchestration requires Claude Agent SDK. This skill provides step design patterns and contracts.

The Four Steps

StepTypeOutputCharacteristic
/planDeterministic outputSpec fileKnown path, structured content
/buildNon-deterministicCode changesCreative, adaptive
/reviewDeterministic outputPASS/FAIL reportStructured, parseable
/fixNon-deterministicTargeted fixesFocused, corrective

Step Design Process

Step 1: Define the Contract

Each step needs a clear contract:

# Plan Step Contract
class PlanContract:
    inputs = {
        "task_description": str,      # What to implement
        "issue_context": str | None,  # Optional context
        "project_path": str           # Where to work
    }
    outputs = {
        "spec_file_path": str,        # Path to generated spec
        "success": bool
    }

Step 2: Ensure Isolation

Each step must be runnable independently:

Isolation Requirements:

  1. Accept inputs via arguments or file paths (not workflow state)
  2. Produce outputs to known, predictable locations
  3. Not depend on other steps having run
  4. Complete or fail cleanly (no partial states)
  5. Be testable in isolation

Step 3: Define Success Criteria

Each step needs clear success indicators:

Plan Step:

  • [ ] Spec file created at expected path
  • [ ] Spec includes acceptance criteria
  • [ ] Scope is bounded and achievable
  • [ ] Technical approach defined

Build Step:

  • [ ] All spec requirements addressed
  • [ ] Code compiles/lints
  • [ ] Tests pass (if exist)
  • [ ] Changes committed

Review Step:

  • [ ] All code reviewed against spec
  • [ ] Issues categorized by severity
  • [ ] Clear PASS/FAIL determination
  • [ ] Actionable issue descriptions

Fix Step:

  • [ ] All flagged issues addressed
  • [ ] No new issues introduced
  • [ ] Changes committed

Step 4: Design Error Handling

Plan failure modes and recovery:

StepFailure ModeRecovery
PlanCan't generate specRetry with more context
BuildBuild errorsReturn to plan
ReviewCritical issuesProceed to fix
FixCan't fixEscalate to human

Step 5: Define Composition Rules

How steps chain together:

plan_build:
  Plan → Build

plan_build_review:
  Plan → Build → Review

plan_build_review_fix:
  Plan → Build → Review → [if FAIL] → Fix → [loop until PASS or max]

Step Templates

Plan Step Template

# /plan

You are a planning agent. Create an implementation specification.

## Inputs

- Task: $ARGUMENTS
- Project: $PROJECT_PATH

## Output

Create a spec file at: specs/{type}-{id}-{description}.md

## Spec Structure

1. **Summary**: One paragraph overview
2. **Requirements**: Numbered list of what must be done
3. **Acceptance Criteria**: How to verify completion
4. **Technical Approach**: How to implement
5. **Files to Modify**: List of affected files

## Success Criteria

- [ ] Spec file created
- [ ] All sections complete
- [ ] Scope is achievable

Build Step Template

# /build

You are an implementation agent. Execute the plan.

## Inputs

- Plan: $ARGUMENTS (path to spec file)
- Project: $PROJECT_PATH

## Process

1. Read the spec file
2. Implement each requirement
3. Run tests if they exist
4. Commit changes with "build: " prefix

## Success Criteria

- [ ] All requirements addressed
- [ ] Code compiles
- [ ] Tests pass
- [ ] Changes committed

Review Step Template

# /review

You are a review agent. Validate the implementation.

## Inputs

- Project: $PROJECT_PATH
- Spec: $SPEC_PATH (optional)

## Output

Structured review report:

STATUS: PASS | FAIL

ISSUES (if FAIL):
- [severity] description

## Severity Levels

- CRITICAL: Must fix before merge
- HIGH: Should fix before merge
- MEDIUM: Consider fixing
- LOW: Optional improvement

## Success Criteria

- [ ] All code reviewed
- [ ] Issues categorized
- [ ] Clear PASS/FAIL status

Fix Step Template

# /fix

You are a fix agent. Resolve review issues.

## Inputs

- Issues: $ARGUMENTS (from review output)
- Project: $PROJECT_PATH

## Process

1. Parse issue list
2. Fix each issue in priority order
3. Verify fix doesn't introduce new issues
4. Commit with "fix: " prefix

## Success Criteria

- [ ] All issues addressed
- [ ] No new issues introduced
- [ ] Changes committed

Loop Limit Pattern

Prevent infinite fix loops:

MAX_FIX_ATTEMPTS = 3

for attempt in range(MAX_FIX_ATTEMPTS):
    review = await run_review()
    if review.status == "PASS":
        break
    await run_fix(review.issues)
else:
    # Max attempts reached
    escalate_to_human("Max fix attempts reached")

Output Format

## Step Design Specification

### Step: [Name]

**Type:** Deterministic Output | Non-Deterministic

**Purpose:** [What this step accomplishes]

### Contract

**Inputs:**
| Name | Type | Required | Description |
| --- | --- | --- | --- |
| [input] | [type] | [yes/no] | [description] |

**Outputs:**
| Name | Type | Description |
| --- | --- | --- |
| [output] | [type] | [description] |

### Success Criteria

- [ ] [Criterion 1]
- [ ] [Criterion 2]
...

### Error Handling

| Failure Mode | Recovery Action |
| --- | --- |
| [mode] | [action] |

### Template

[Slash command template]


### Composition Rules

[How this step chains with others]

Design Checklist

  • [ ] Contract inputs defined
  • [ ] Contract outputs defined
  • [ ] Isolation verified (standalone runnable)
  • [ ] Success criteria specified
  • [ ] Error handling planned
  • [ ] Composition rules documented
  • [ ] Template created
  • [ ] Tests designed

Anti-Patterns

Anti-PatternProblemSolution
Coupled stepsCan't run independentlyPass data via files
Unbounded loopsInfinite fix attemptsMAX_FIX_ATTEMPTS
Undefined contractsUnpredictable I/OExplicit contracts
Missing success criteriaCan't verify completionDefine criteria
No error handlingSilent failuresPlan recovery actions

Cross-References

  • @composable-steps.md - Step memory file
  • @composable-primitives.md - General primitives
  • @adw-framework.md - ADW overview
  • @template-engineering.md - Template patterns

Version History

  • v1.0.0 (2026-01-01): Initial release (Lesson 14)

Last Updated

Date: 2026-01-01 Model: claude-opus-4-5-20251101

平台分布

Antigravity

26.38%
按下载量换算26

windsurf

21.41%
按下载量换算21

trae

17.25%
按下载量换算17

Claude Code

13.08%
按下载量换算13

OpenCode

7.02%
按下载量换算7

Gemini CLI

3.72%
按下载量换算4

安全审计

Gen Agent Trust Hub

通过

Socket

通过

Snyk

通过

权限和风险

只读

该 Skill 主要提供规则、说明或参考内容,本身偏只读;真正读写文件、联网或执行命令仍取决于宿主 Agent 的任务。

安装前确认

本站仅展示第三方公开信息,不托管安装包,不提供自动安装或运行环境。安装前应自行审查源码、依赖和命令行为。

继续浏览同类 Skills