Create Demo

Create Demo

Create a standalone, minimal demo that reproduces a single core mechanism from an analyzed AI agent. Supports both coding agent mechanisms (D2-D7) and platform mechanisms (D9-D12).

Trigger

/create-demo <agent-name> <mechanism>

Example: /create-demo aider repomap Example: /create-demo openclaw channel-router

Prerequisites

1. Analysis doc must exist at docs/<agent-name>.md (run /analyze-agent first)
2. The mechanism must be identified in the analysis doc (D2-D7 for coding, D9-D12 for platform)

Workflow

Step 1: Read Analysis

Read docs/<agent-name>.md and identify the target mechanism:

• What problem does it solve?
• What's the core algorithm/flow?
• What are the key data structures?
• What dependencies does the original use?

Step 2: Design Simplification

Apply these simplification principles:

• Single mechanism: Focus on ONE mechanism only, remove all others
• Minimal dependencies: Use the fewest possible libraries
• No infrastructure: Remove git integration, database caching, CLI frameworks, config systems
• Hardcode when possible: Replace configurable options with sensible defaults
• 语言选择: 默认 Python。当机制依赖语言特性时用原生语言，README 须含 "为何选择此语言" 节
• Simple I/O: Print to stdout, no rich formatting or interactive prompts
• Inline over import: If a helper is < 20 lines, inline it rather than importing from another demo

Step 3: Create File Structure

Create demos/<agent-name>/<mechanism>/ with the structure matching the chosen language:

Python:

demos/<agent-name>/<mechanism>/
├── README.md           # From template, filled in
├── requirements.txt    # Only if external deps needed (not for stdlib-only demos)
├── main.py             # Entry point (or split into logical modules)
└── sample_*/           # Sample data/project if needed for demonstration

TypeScript:

demos/<agent-name>/<mechanism>/
├── README.md           # From template, filled in
├── package.json        # Dependencies
├── tsconfig.json       # TypeScript config
├── main.ts             # Entry point
└── sample_*/           # Sample data/project if needed for demonstration

Rust:

demos/<agent-name>/<mechanism>/
├── README.md           # From template, filled in
├── Cargo.toml          # Project config & dependencies
├── src/
│   └── main.rs         # Entry point
└── sample_*/           # Sample data/project if needed for demonstration

Step 4: Write README

Use the template from demos/TEMPLATE/README.md and fill in:

• Goal: One sentence describing what this demo reproduces
• 原理: How the mechanism works in the original agent (2-3 paragraphs)
• 运行: Exact commands to run the demo
• 文件结构: Actual file tree
• 关键代码解读: Core algorithm with inline comments
• 与原实现的差异: What was simplified and what was preserved
• 相关文档: Must include 基于 commit: <short-sha> and 核心源码: <原项目中对应的文件路径> (the source path is used by /check-updates to match changes)

Step 5: Implement Core Logic

Guidelines:

• Target ~100-200 lines of core logic per demo
• LLM 调用按语言选择对应库和运行方式：

• Each demo must be independently runnable
• Include sample data/projects so the demo works out of the box
• Add clear print statements showing what's happening at each step

Step 5.5: Complete Integration Demo (mini-agent)

When all MVP component demos (and platform mechanism demos if applicable) for an agent are complete, create a combined integration demo:

• 目录: demos/<agent>/mini-<agent>/
• 目标: 串联所有 MVP 组件（+ 平台机制 demo，如适用）为一个最简可运行 agent
• 核心原则: 必须 import 兄弟 MVP/平台 demo 的模块，而非把所有代码重写到一个文件中。mini-agent 本身只实现 Core Loop 逻辑，其余能力全部通过 import 复用
• 行数: mini-agent 自身 100-200 行（不含导入的模块）
• 前提: 所有 MVP 组件 demo 已完成；平台型 agent 的平台机制 demo 也需完成
• 语言: 与多数 MVP 组件一致（混合则用 Python）
• 平台型 agent 特别说明: 串联 demo 应体现完整的消息流路径——从通道接入到 agent 执行再到响应返回，而非只串联编码部分

模块提取约定:

• 每个 MVP demo 应将核心逻辑提取到独立模块文件（如 prompt-assembly/assembler.py）
• main.py 保留 demo 演示代码，从模块 import 核心类/函数
• mini-agent 通过 sys.path 添加兄弟目录来 import 这些模块

示例（aider）:

# mini-aider/main.py
from assembler import PromptAssembler       # ← prompt-assembly/
from parser import find_edit_blocks         # ← search-replace/
from replacer import apply_edit             # ← search-replace/
from parsers import generate_reflection     # ← llm-response-parsing/

Step 6: Verify

Run the demo and confirm:

• It executes without errors
• It demonstrates the core mechanism clearly
• Output is understandable to someone learning the mechanism

Code Style

• Use type hints for function signatures
• Add docstrings for the main functions explaining the mechanism
• Keep classes minimal — prefer functions unless state management is essential
• Name variables to match the original codebase's terminology where possible

Output

The completed demo directory at demos/<agent-name>/<mechanism>/.

After Creating Demo

1. Update the agent's overview at demos/<agent-name>/README.md: mark the mechanism as [x] and update the progress line. Format:
   • Coding agents: MVP: X/N | 进阶: Y/M | 串联: Z/1 | 总计: A/K
   • Platform agents: MVP: X/N | 平台: P/Q | 进阶: Y/M | 串联: Z/1 | 总计: A/K
2. Update agents.yaml status if needed (e.g. pending → in-progress)
3. Run npm run progress to update the CLAUDE.md progress section (or it will auto-update on next commit)
4. Confirm the demo README's "相关文档" section includes 基于 commit: <short-sha> (read from agents.yaml analyzed_commit). This provides a baseline for /check-updates demo impact assessment.