agentDevCoder

Use this skill in the scenario of intelligent agent application development.

16 stars

Best use case

agentDevCoder is best used when you need a repeatable AI agent workflow instead of a one-off prompt.

Use this skill in the scenario of intelligent agent application development.

Teams using agentDevCoder should expect a more consistent output, faster repeated execution, less prompt rewriting.

When to use this skill

  • You want a reusable workflow that can be run more than once with consistent structure.

When not to use this skill

  • You only need a quick one-off answer and do not need a reusable workflow.
  • You cannot install or maintain the underlying files, dependencies, or repository context.

Installation

Claude Code / Cursor / Codex

$curl -o ~/.claude/skills/agentdevcoder/SKILL.md --create-dirs "https://raw.githubusercontent.com/diegosouzapw/awesome-omni-skill/main/skills/ai-agents/agentdevcoder/SKILL.md"

Manual Installation

  1. Download SKILL.md from GitHub
  2. Place it in .claude/skills/agentdevcoder/SKILL.md inside your project
  3. Restart your AI agent — it will auto-discover the skill

How agentDevCoder Compares

Feature / AgentagentDevCoderStandard Approach
Platform SupportNot specifiedLimited / Varies
Context Awareness High Baseline
Installation ComplexityUnknownN/A

Frequently Asked Questions

What does this skill do?

Use this skill in the scenario of intelligent agent application development.

Where can I find the source code?

You can find the source code on GitHub using the link provided at the top of the page.

Related Guides

SKILL.md Source

# SKILL: 智能体应用开发(Agent App Development)

> 目标:把一个“能用、可控、可扩展”的智能体应用从想法落地到工程实现,覆盖:工具调用、检索、记忆、工作流编排、评测与上线。

---

## 1. 适用范围与不适用范围

### ✅ 适用范围

- 单/多智能体应用架构设计(planner / executor / critic / router)
- 工具系统:function calling、MCP、HTTP 工具、数据库工具
- 工作流编排:LangGraph / 状态机 / plan-act-observe-replan
- 检索增强:RAG(向量库/混合检索/重排序/引用与溯源)
- 记忆:短期/长期/用户偏好、结构化存储(SQL/Key-Value/Docstore)
- 可靠性:权限约束、人类确认(HITL)、回退策略、重试与幂等
- 评测:离线回放、golden set、自动评审、成本/延迟/成功率指标
- 工程化:日志/追踪、配置化、部署(容器化/无服务器/队列)

### ❌ 不适用范围

- 要求绕过安全限制、窃取隐私、未授权访问(拒绝)
- “完全自动化替代决策”且不允许人工确认的高风险场景
- 无可测目标的泛泛需求(会先给默认目标与可度量指标)

---

## 2. 交付物(Outputs)

- 架构图(文字版组件与数据流)
- 状态定义(State Schema)与节点职责
- 工具清单(Tool Contract:name/args/return/error)
- 提示词模板(System/Developer/Node Prompts)
- 评测方案与指标表
- 最小可运行骨架(伪代码或可粘贴代码片段)

---

## 3. 输入格式(Inputs)

**最小输入**:

- 用例:用户要完成什么任务(例如“读论文 → 写综述 → 生成表格”)
- 资源:可用的数据源/工具(本地文件、API、数据库、网页)
- 约束:成本、时延、离线/在线、是否需要引用来源

**推荐输入**:

- 失败样例与成功样例各 3 个
- 需要的工具权限边界(读/写/删除/支付等)
- 期望的 UI/交互形态(聊天、表单、批处理、插件)

---

## 4. 设计原则(Agent Skill 核心)

1. **可控优先**:先把边界与权限写清楚,再追求自动化
2. **可测优先**:每个节点要有可度量的输入输出与成功条件
3. **最小闭环**:先做能跑通的 end-to-end MVP,再逐步增强
4. **显式状态**:用结构化 State 传递信息,避免“凭空记忆”
5. **工具契约**:工具返回统一结构,错误可恢复(recoverable)
6. **失败即路径**:为常见失败设定回退策略与用户澄清模板

---

## 5. 标准工作流(Workflow)

### Step 0:定义目标与评测

- 目标:成功率、准确率、引用覆盖率、成本、延迟
- 评测集:golden queries(最少 20 条)+ 边界案例(最少 10 条)

### Step 1:拆解为节点(Nodes)

建议 4~7 个节点,典型如下:

- `Intake`:收集意图/约束/上下文,形成结构化需求
- `Plan`:生成可执行计划(含工具调用点与停止条件)
- `Act`:执行工具调用(幂等、重试、限流)
- `Observe`:解析工具结果,写入 State
- `Judge`:质量审查(引用、格式、覆盖、风险)
- `Respond`:面向用户生成最终输出
- `Repair`:失败修复(缩小范围、换策略、请求澄清)

### Step 2:定义 State Schema

- `goal`:用户目标
- `constraints`:成本/时延/安全
- `artifacts`:中间产物(摘要、表格、代码片段)
- `tool_calls`:历史工具调用(request/response/error)
- `citations`:引用与出处
- `decision_log`:关键决策理由(用于调试与复盘)

### Step 3:工具契约(Tool Contract)

统一返回结构(示例):

- `ok: bool`
- `data: any`
- `error: {type, message, retryable, details}`
- `meta: {latency_ms, cost, source}`

### Step 4:失败处理与回退

- 工具失败:指数退避 + 降级模式(少工具/少步骤)
- 检索失败:扩大 query、改用混合检索、请求用户提供关键词
- 输出失败:缩短答案、分段输出、先给结论后给依据

---

## 6. 提示词与路由模板

### 6.1 节点提示词结构

- Role:你是谁、你的职责
- Inputs:你从 State 读什么
- Outputs:你必须写回什么字段
- Constraints:不能做什么、必须引用什么
- Rubric:通过/失败标准

### 6.2 路由(Router)规则

- 依据:intent、所需工具、风险等级、是否需要澄清
- 输出:下一节点 + 理由(写入 decision_log)

---

## 7. RAG 与记忆(推荐实现)

### 7.1 RAG 最小形态

- chunk → embedding → top-k → 生成(带引用)

### 7.2 增强形态

- 混合检索(BM25 + 向量)
- 重排序(cross-encoder / reranker)
- 引用对齐(回答段落 ↔ chunk id)

### 7.3 记忆分层

- 短期:对话窗口内 state
- 长期:用户偏好/项目事实(结构化表)
- 可撤销:用户可要求遗忘

---

## 8. 评测与可观测性(Observability)

### 8.1 指标

- 成功率:任务完成 / 总任务
- 工具成功率:ok / 调用
- 引用覆盖率:含引用回答 / 需要引用回答
- 幻觉率:错误事实 / 抽检
- 成本:token + 工具成本
- 延迟:p50/p95

### 8.2 日志与追踪

- trace_id / run_id
- node_latency、tool_latency
- state_diff(关键字段变化)
- error 分类与频次

---

## 9. 反例(Anti-patterns)

- 把所有逻辑塞进一个超长 prompt,没有显式 state
- 工具返回格式混乱,错误不可恢复
- 没有评测集,靠主观感觉迭代
- 没有权限边界与用户确认,高风险动作直接执行

---

## 10. 交互方式(How to Ask Me)

你可以这样给我需求:

### A. 做一个可落地的智能体

- 目标:帮我把论文 PDF → 自动提炼相关工作 → 输出表格
- 工具:本地 PDF、向量库、网页搜索(可选)
- 约束:需要引用来源;成本尽量低
- 输出:Markdown + 可复制表格

### B. 诊断智能体失败

- 失败案例:贴出用户输入、工具返回、最终输出
- 现象:引用丢失/走错路由/循环调用
- 期望:修复策略 + 节点/状态改造建议

---

## 11. 质量标准(Definition of Done)

- ✅ MVP 可跑通:给定 20 条测试集成功率达标
- ✅ 输出稳定:同输入多次运行差异可解释
- ✅ 可控:高风险动作默认需要确认
- ✅ 可观测:能定位“卡在哪个节点/哪个工具”
- ✅ 可扩展:新增工具/新增节点不破坏整体

Related Skills

bgo

10
from diegosouzapw/awesome-omni-skill

Automates the complete Blender build-go workflow, from building and packaging your extension/add-on to removing old versions, installing, enabling, and launching Blender for quick testing and iteration.

Coding & Development

mcp-create-declarative-agent

16
from diegosouzapw/awesome-omni-skill

Skill converted from mcp-create-declarative-agent.prompt.md

MCP Architecture Expert

16
from diegosouzapw/awesome-omni-skill

Design and implement Model Context Protocol servers for standardized AI-to-data integration with resources, tools, prompts, and security best practices

mathem-shopping

16
from diegosouzapw/awesome-omni-skill

Automatiserar att logga in på Mathem.se, söka och lägga till varor från en lista eller recept, hantera ersättningar enligt policy och reservera leveranstid, men lämnar varukorgen redo för manuell checkout.

math-modeling

16
from diegosouzapw/awesome-omni-skill

本技能应在用户要求"数学建模"、"建模比赛"、"数模论文"、"数学建模竞赛"、"建模分析"、"建模求解"或提及数学建模相关任务时使用。适用于全国大学生数学建模竞赛(CUMCM)、美国大学生数学建模竞赛(MCM/ICM)等各类数学建模比赛。

matchms

16
from diegosouzapw/awesome-omni-skill

Mass spectrometry analysis. Process mzML/MGF/MSP, spectral similarity (cosine, modified cosine), metadata harmonization, compound ID, for metabolomics and MS data processing.

managing-traefik

16
from diegosouzapw/awesome-omni-skill

Manages Traefik reverse proxy for local development. Use when routing domains to local services, configuring CORS, checking service health, or debugging connectivity issues.

managing-skills

16
from diegosouzapw/awesome-omni-skill

Install, find, update, and manage agent skills. Use when the user wants to add a new skill, search for skills that do something, check if skills are up to date, or update existing skills. Triggers on: install skill, add skill, get skill, find skill, search skill, update skill, check skills, list skills.

manage-agents

16
from diegosouzapw/awesome-omni-skill

Create, modify, and manage Claude Code subagents with specialized expertise. Use when you need to "work with agents", "create an agent", "modify an agent", "set up a specialist", "I need an agent for [task]", or "agent to handle [domain]". Covers agent file format, YAML frontmatter, system prompts, tool restrictions, MCP integration, model selection, and testing.

maintainx-automation

16
from diegosouzapw/awesome-omni-skill

Automate Maintainx tasks via Rube MCP (Composio). Always search tools first for current schemas.

mailsoftly-automation

16
from diegosouzapw/awesome-omni-skill

Automate Mailsoftly tasks via Rube MCP (Composio). Always search tools first for current schemas.

mails-so-automation

16
from diegosouzapw/awesome-omni-skill

Automate Mails So tasks via Rube MCP (Composio). Always search tools first for current schemas.