Token导航 LogoToken导航TokenDH.com
研究检索需要联网github未标认证来源可访问许可证需确认审计未展示

agent%3apromptAgent 3aprompt 搜索

Agent Skill

用于辅助提示词、系统指令、Agent 行为约束和工作流模板的整理。它适合让 Agent 规范任务边界、统一输出格式、拆分操作步骤或优化提示词可复用性。使用时需要保留真实业务约束,不要把示例当硬规则;涉及自动执行、外部工具或高风险操作时,应在提示词中明确确认步骤、权限边界和失败处理方式。

总安装

188

周安装

8

GitHub Stars

公开资料未说明

下载量

66
CodexClaudeCursorGemini CLI

安装说明

本站只整理中文说明和来源信息,不托管安装包,也不代用户安装。

GitHub

来源数

2

许可证

unknown

最后核验

2026-05-01

来源状态

来源可访问

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:agent%3aprompt(Agent 3aprompt 搜索)
来源仓库:https://github.com/ikatsuba/skills
仓库路径:skills/agent%3Aprompt
安装命令:
npx skills add https://github.com/ikatsuba/skills --skill agent:prompt
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。该命令会通过 npx skills 从第三方来源获取 Skill;本站只展示命令,不托管安装包,也不自动执行。

skills.shnpx skills
npx skills add https://github.com/ikatsuba/skills --skill agent:prompt

简介

用于辅助提示词与系统指令的设计优化,提升 Agent 行为一致性。

  • 适合在 Codex、Claude、Cursor、Gemini CLI 中规范任务边界与输出格式。
  • 基于种子提示法构建可复用模板,支持少样本示例与生产级优化。
  • 使用时需保留真实业务约束,避免将示例当作硬规则。
  • agent%3aprompt 属于研究检索类 Skill,可作为该场景下的辅助能力补充。

SKILL.md

Prompt Engineering

Guides the user through designing and optimizing prompts for AI agents. Based on "Principles of Building AI Agents" (Bhagwat & Gienow, 2025), Part I: Prompting a Large Language Model.

When to use

Use this skill when the user needs to:

  • Choose the right model and provider for their agent
  • Write or improve system prompts and agent instructions
  • Design few-shot examples for consistent output
  • Optimize prompts for production (cost, quality, latency)
  • Bootstrap a prompt from scratch using the seed crystal approach

Instructions

Step 1: Understand the Use Case

Use the AskUserQuestion tool to gather context:

  1. What does the agent do? (task type: classification, generation, extraction, conversation, code, reasoning)
  2. What is the expected input format? (free text, structured data, images, code)
  3. What is the expected output format? (free text, JSON, code, decision)
  4. What quality bar is needed? (prototype, internal tool, customer-facing, high-stakes)
  5. What are the constraints? (latency, cost, context size)

Step 2: Model Selection

Guide the user through model choice. Apply the principle: start expensive, optimize later.

Use AskUserQuestion to present options:

## Model Selection Matrix

| Factor | Small/Fast | Medium | Large/Capable |
|--------|-----------|--------|---------------|
| **Use case** | Classification, routing, simple extraction | Conversation, summarization, tool calling | Complex reasoning, code gen, multi-step planning |
| **Latency** | < 500ms | 1-3s | 3-10s |
| **Cost** | $0.25-1/M tokens | $3-5/M tokens | $15-75/M tokens |
| **Examples** | Haiku, GPT-4o-mini, Gemini Flash | Sonnet, GPT-4o, Gemini Pro | Opus, o1/o3, Gemini Ultra |

Key principles:

  • Start hosted — use cloud APIs (Anthropic, OpenAI, Google) even if you plan open-source later
  • Use model routing — abstract the provider so you can swap models without rewriting code
  • Consider context windows — if your use case needs large input (codebases, long documents), pick models with large context windows
  • Reasoning models — best for complex analysis but need lots of context upfront (many-shot prompting); without good context they go off the rails

Output a recommendation:

## Recommended Model

**Primary:** [Model] — [Why]
**Fallback:** [Model] — [For cost optimization / lower-stakes tasks]
**Context window:** [Size] — [Sufficient for X pages / Y tokens of input]

Step 3: Prompt Architecture

Design the prompt structure. Every production prompt has these layers:

## Prompt Architecture

### 1. System Prompt
[Agent identity, role, tone, constraints, persona]
- Sets characteristics and behavior boundaries
- Good for shaping tone; usually does not improve accuracy alone

### 2. Context Block
[Data the agent needs to do its job]
- Retrieved documents (RAG)
- User profile / session state
- Tool descriptions and schemas
- Use XML-like tags to structure: <context>, <user_data>, <documents>

### 3. Task Instructions
[What to do with the context]
- Step-by-step instructions
- Output format specification
- Edge case handling

### 4. Examples (Few-Shot)
[Input/output pairs showing desired behavior]
- Zero-shot: no examples (simplest, least control)
- Single-shot: one example (establishes format)
- Few-shot: 3-5 examples (most control, highest quality)

### 5. User Input
[The actual user message / query]

Step 4: Write the System Prompt

Guide the user through writing each section. Use AskUserQuestion at each step.

Seed Crystal Approach: If the user is starting from scratch, offer to bootstrap:

  1. Describe the agent's purpose in 2-3 sentences
  2. Ask the target LLM to generate a v1 system prompt
  3. Refine iteratively

Formatting principles:

  • CAPITALIZATION adds weight to important words (use sparingly)
  • XML-like tags help models parse structure: <task>, <context>, <constraints>, <examples>
  • Be extremely specific — production prompts are long and detailed
  • Include what the agent should NOT do (prohibitions prevent common failures)

Output a complete system prompt:

## System Prompt

<system>
You are [role/identity].

## Your Task
[Clear description of what to do]

## Constraints
- [Constraint 1]
- [Constraint 2]
- NEVER [prohibition]

## Output Format
[Exact format specification]

## Examples

<example>
<input>[Example input]</input>
<output>[Example output]</output>
</example>

<example>
<input>[Another input]</input>
<output>[Another output]</output>
</example>
</system>

Step 5: Design Few-Shot Examples

Help the user create high-quality examples. Use AskUserQuestion to gather real scenarios.

Guidelines:

  • Cover the most common case first
  • Include at least one edge case
  • Show the exact output format you expect
  • If the agent handles errors, show an error example
  • 3-5 examples is the sweet spot (more = more control but higher cost)
## Few-Shot Examples

| # | Input | Expected Output | Covers |
|---|-------|----------------|--------|
| 1 | [Common case] | [Ideal response] | Happy path |
| 2 | [Edge case] | [Correct handling] | Boundary condition |
| 3 | [Ambiguous input] | [Clarification response] | Uncertainty handling |
| 4 | [Error case] | [Graceful failure] | Error path |

Step 6: Production Optimization

Review the prompt for production readiness:

## Production Checklist

### Quality
- [ ] System prompt clearly defines agent identity and boundaries
- [ ] At least 3 few-shot examples covering common + edge cases
- [ ] Output format is unambiguous (JSON schema, structured template)
- [ ] Prohibitions prevent known failure modes
- [ ] Formatting uses XML tags for structure

### Cost
- [ ] Prompt tokens are minimized (no redundant instructions)
- [ ] Examples are concise but representative
- [ ] Context is filtered to what's relevant (not dumping everything)
- [ ] Consider using a smaller model for simple subtasks

### Latency
- [ ] Streaming is enabled for user-facing responses
- [ ] Long prompts are cached where supported
- [ ] Model size matches latency requirements

### Robustness
- [ ] Tested with adversarial inputs
- [ ] Tested with empty/minimal input
- [ ] Tested with very long input (near context limit)
- [ ] Output format holds across model temperature settings

Step 7: Summarize and Offer Next Steps

Present all findings to the user as a structured summary in the conversation. Do NOT write to .specs/ — this skill works directly.

Use AskUserQuestion to offer:

  1. Test the prompt — try it with sample inputs right now
  2. Write the prompt to a file — if the user wants to save it to their codebase
  3. Comprehensive design — run agent:design to cover all areas with a spec

Arguments

  • $ARGUMENTS ($0) - Optional description of the agent or path to existing prompt file

Examples:

  • agent:prompt customer-support chatbot — design prompts for a support chatbot
  • agent:prompt src/agents/writer.ts — improve prompts in an existing agent file
  • agent:prompt — start fresh

适合场景

01

用户想查找某类 Agent Skill 时

02

需要根据任务场景推荐可安装能力包时

03

需要对比不同来源的安装命令和来源信息时

能力概览

能力 1

按任务关键词查找相关 Skills

能力 2

展示可复制的安装命令

能力 3

保留来源站点、仓库和原始说明,方便继续核验

安装后应在对应宿主中按原始 README 的触发条件使用;具体调用方式请以来源页面和 README 为准。

平台分布

Codex

35.15%
按下载量换算23

Claude

29.89%
按下载量换算20

Cursor

18.49%
按下载量换算12

Gemini CLI

10.12%
按下载量换算7

安全审计

暂无安全审计结果可展示。

权限和风险

需要联网

该 Skill 可能需要联网访问来源站点、仓库或外部 API;具体网络访问范围需要结合源码和 README 复核。

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。当前只有一个来源,正式发布前建议补源仓库或其他目录站核验。

来源信息

继续浏览同类 Skills