This is an internal workshop that shows how AI agentic coding practices improve product development at Crescendo Lab (CL) through a real-world legacy system modernization case study.
By applying AI agentic coding practices to existing brownfield codebases, our teams delivered 1.6x faster compared to original AI auto-complete approach while significantly improving test quality.
By the end of this workshop, you'll learn:
- Context-Driven Development - Provide comprehensive context (PRD, KB, domain models) to AI for better code generation
- Design-First Approach - Plan interfaces and tasks before implementation
- Knowledge Base Management - Extract and maintain domain expertise for better AI context
Requirements: AI coding assistant (Cursor, Claude, etc.) + basic programming experience
License: Apache 2.0 - free to use and adapt
How to Use This README:
- AI Coding Principles - Learn the 3 C's framework and AI workflow (read this first!)
- Workshop Outline - See all 4 tutorials with time estimates
- Learning Path: Tutorial 1 (90min) → Tutorial 2 (40min) → Tutorial 3 (40min) → Tutorial 4 (20min)
- Stuck? Check FAQ sections in tutorials. Confusion is normal and part of the learning!
- Case Study - Understand what you're building (don't worry if it's complex!)
- Codebase Outline - Find working directories and reference materials
- Workshop Setup - Detailed environment setup
At Crescendo Lab, we follow the 3 C's of AI coding:
- Context: Provide comprehensive background (PRD, KB, design docs) before coding
- Control: Humans drive decisions, AI executes
- Critique: Always validate and verify outputs. No vibe coding.
- New Spec (User Story + Use Cases) - Requirements for new features
- Existing Spec - Current system documentation and requirements
- Engineer Knowledge - Domain expertise and tribal knowledge
- Codebase - Current implementation and architecture
- Copilot: Human-guided AI assistance for planning, clarification, and review
- Agentic: Autonomous AI execution for implementation and testing
- Extract knowledge (Copilot) → KB (*1)
- Convert engineer knowledge and existing code into structured knowledge base
- Surface in AI tools (Cursor, Claude, etc.)
- Make context available to AI coding assistants
- Clarify requirements (Copilot) (*1)
- AI asks clarifying questions to ensure complete understanding
- Design domain & interface (Copilot)
- Human-driven architectural decisions and API design
- AI implements features based on plans (Agentic) (*2)
- Autonomous implementation with error correction
- AI writes tests to validate functionality, based on use cases (Agentic) (*2)
- Comprehensive test coverage derived from the requirements
- Review & Refactor (Copilot)
- Update KB (Copilot)
- Commit
Feedback Loops
- (*1): AI Q&A loop: Iterative clarification and knowledge extraction
- (*2): AI self-correction loop: Automatic refinement during implementation and testing
- Follow agentic coding flow
- Build new functions and flows alongside old ones
- Progressive rollout with feature flags
- Legacy stays live during transition
- Quick rollback if issues occur
This workshop guides participants through step-by-step agentic coding practice using the Omnichannel Auto-Reply case study across the following sections:
| Section | Topic | Description | Estimated Time | Tutorial |
|---|---|---|---|---|
| 1 | Rewrite Brownfield | Build Omnichannel Auto-Reply trigger logic with agentic coding | 1h 20m | link |
| 2 | Extend Functions | Extend Omnichannel Auto-Reply to support IG story-specific | 40m | link |
| 3 | KB Extraction | Extract knowledge from codebase and tribal knowledge for critical features | 40m | link |
| 4 | Sharpen your prompts | Patterns for effective agent instructions | 20m | link1, link2 |
Auto-Reply is a critical product feature at CL that helps our customers automatically respond to their end users' messages using pre-defined keywords and time schedules via their LINE official account (LINE OA).
Currently, CL is planning to extend Auto-Reply to Facebook Messenger and Instagram accounts, based on the existing LINE version, called Omnichannel Auto-Reply.
The following diagram illustrates the auto-reply processing flow from webhook reception to message delivery:
graph TD
A["Incoming Message<br/>(LINE/FB/IG Webhook)"] --> B["Webhook Event Processing"]
B --> C["Extract Message Content"]
C --> D["Load Org/Channel Info &<br/> Auto-Reply Rules"]
D --> E["Trigger Validation"]
E --> F{"Valid Trigger<br/>Found?"}
F -->|Yes| G["Build Reply Message"]
F -->|No| J["No Response"]
G --> I["Send Reply Message<br/>via Channel API"]
I --> K["End"]
J --> K
style A fill:#e1f5fe
style E fill:#fff3e0
style G fill:#c8e6c9
style I fill:#c8e6c9
style J fill:#ffcdd2
The key differences between the new and existing Auto-Reply:
| Feature | Legacy System | Omnichannel System |
|---|---|---|
| Platform Support | LINE only | LINE, Facebook Messenger, Instagram DMs |
| Architecture | Single-channel, direct access | Multi-channel with parent-child model |
| Data Model | WebhookTriggerSetting → WebhookTriggerMessage | AutoReply → AutoReplyChannelSetting → AutoReplyMessage |
| Keyword Triggers | ✅ MESSAGE | ✅ MESSAGE |
| Keyword Match | Case-sensitive exact match | Case-insensitive + trim spaces exact match |
| Multiple Keywords per Rule | ❌ One trigger_code per WebhookTriggerSetting | ✅ Multiple keywords per AutoReply rule |
| General Time-based Triggers | ✅ Schedule type priority | ✅ Schedule type priority |
| Cross-Platform Rules | ❌ LINE only | ✅ One rule → multiple channels |
| IG Story Integration | ❌ Not supported | ✅ 4-level priority system with story-specific triggers |
| Conflict Resolution | First match wins, manual admin avoidance | ✅ Automatic constraint validation |
| Cross-Channel Reporting | ❌ LINE only | ✅ Channel-specific breakdowns |
| Rule Management | Per-bot configuration | ✅ Hierarchical parent-child model |
| Backward Compatibility | N/A | ✅ Full LINE compatibility maintained |
Videos
Omnichannel Auto-Reply (Keyword + General):
Auto-Reply.Keyword.+.General.mp4
Omnichannel Auto-Reply (IG Story-Specific):
Auto-Reply.IG-specific.mp4
Help Center Documentation
PRDs
KB & Domain models
This workshop provides both Go and Python workspaces and reference materials:
Working Directories
- go_src/ - Go implementation workspace for hands-on practice
- python_src/ - Python implementation workspace for hands-on practice
Reference Materials
- cheat_sheet/ - Sample implementations and solutions
go/1_rewrite_brownfield/- Go solution for Section 1go/2_extend_feature/- Go solution for Section 2python/1_rewrite_brownfield/- Python solution for Section 1python/2_extend_feature/- Python solution for Section 2
- legacy/ - Original LINE-only implementation for reference
kb/- Knowledge base documentsline/- Legacy Python LINE implementationorganization/- Legacy organization models
- spec/ - Product requirements and specifications
prd-part1.md- Keyword + General Auto-Reply requirementsprd-part2.md- IG Story-specific Auto-Reply requirementsig_story.json- IG Story webhook event sample
Before starting the workshop, ensure you have the following tools installed:
- brew - Package manager for macOS
- Go 1.23+ - Required for the Go workspace
Python Workspace Setup:
cd python_src/
make init
make fmt # Format code
make test # Verify setupGo Workspace Setup:
cd go_src/
make test # Verify setupThis workshop is licensed under Apache 2.0 - all materials in this repository are open source and free to use.
You are encouraged to:
- 🍴 Fork this repository for your own learning and experimentation
- 🏢 Use and adapt these materials in your organization and teams
- 📚 Reference the approaches and techniques in your daily work
- 🤝 Share with colleagues and the broader development community
- 💡 Contribute improvements, fixes, or additional examples back to the project
Minimal requirements - simply preserve copyright notices and mention any changes you make. Feel free to adapt the tutorials, modify the examples, and make them work for your specific use cases. The goal is to help as many developers as possible learn effective AI agentic coding practices.
See LICENSE for full details.