chatbot-builder

Chatbot Builder — Chatbot Construction Pipeline

An agent team collaborates to build a chatbot through persona design > conversation design > NLU > integration > testing.

Execution Mode

Agent Team — Five agents communicate directly via SendMessage and perform cross-validation.

Agent Composition

Agent	File	Role	Type
persona-architect	.claude/agents/persona-architect.md	Bot persona design	general-purpose
conversation-designer	.claude/agents/conversation-designer.md	Conversation scenario design	general-purpose
nlu-developer	.claude/agents/nlu-developer.md	NLU pipeline implementation	general-purpose
integration-engineer	.claude/agents/integration-engineer.md	Channel integration, deployment	general-purpose
dialog-tester	.claude/agents/dialog-tester.md	Quality verification, testing	general-purpose

Workflow

Phase 1: Preparation (performed directly by the orchestrator)

1. Extract the following from user input:
   • Chatbot purpose: Customer service/FAQ/reservations/orders/consultation, etc.
   • Target users: Age, digital literacy, expectation level
   • Integration channels: Slack/KakaoTalk/Telegram/Web/multi-channel
   • Domain knowledge: Business rules, FAQ list, product information, etc.
   • Constraints (optional): Technology stack, response time requirements
2. Create the _workspace/ directory at the project root
3. Organize the input and save it to _workspace/00_input.md
4. Create the _workspace/src/ directory
5. If pre-existing files are available, copy them to _workspace/ and skip the corresponding phase
6. Determine the execution mode based on the scope of the request (see "Execution Modes by Request Scope" below)

Phase 2: Team Assembly and Execution

Order	Task	Owner	Dependencies	Deliverable
1	Persona design	persona	None	_workspace/01_persona_spec.md
2	Conversation design	designer	Task 1	_workspace/02_conversation_design.md
3a	NLU implementation	nlu-dev	Task 2	_workspace/03_nlu_config.md + src/
3b	Integration design	integrator	Task 2	_workspace/04_integration_spec.md + src/
4	Testing	tester	Tasks 3a, 3b	_workspace/05_test_report.md

Tasks 3a (NLU) and 3b (integration) run in parallel.

Inter-agent communication flow:

• persona completes > passes tone and manner guide to designer, passes domain keywords to nlu-dev
• designer completes > passes intent/entity catalog to nlu-dev, passes external integration flows to integrator
• nlu-dev completes > passes NLU interface to integrator, passes test data to tester
• integrator completes > passes test environment info to tester
• tester cross-validates all deliverables. On CRITICAL findings, requests corrections from the relevant agent > rework > re-verification (up to 2 rounds)

Phase 3: Integration and Final Deliverables

Finalize deliverables based on the tester's report:

1. Verify all files in _workspace/ and src/ code
2. Confirm that all CRITICAL findings have been resolved
3. Report the final summary to the user

Execution Modes by Request Scope

User Request Pattern	Execution Mode	Agents Deployed
"Build me a chatbot", "full build"	Full pipeline	All 5 agents
"Just do the conversation design", "write scenarios"	Design mode	persona + designer
"Just develop the NLU" (design complete)	NLU mode	nlu-dev + tester
"Integrate chatbot with KakaoTalk" (implementation complete)	Integration mode	integrator + tester
"Test the chatbot" (implementation complete)	Test mode	tester only

Reusing existing files: If the user provides an existing conversation design document or NLU configuration, copy those files to _workspace/ and skip the corresponding steps.

Data Transfer Protocol

Strategy	Method	Purpose
File-based	_workspace/ directory	Design documents and configuration sharing
Message-based	SendMessage	Real-time key information transfer, correction requests
Code-based	_workspace/src/	Executable source code

Error Handling

Error Type	Strategy
Insufficient domain knowledge	Request additional FAQ list/business information from user, supplement with web search
Channel API changes	Verify latest API documentation via WebFetch, then update integration code
NLU accuracy below threshold	Augment training data > redesign prompts > strengthen fallback, in that order
Agent failure	Retry once > if still failing, proceed without that deliverable and note in report

Test Scenarios

Normal Flow

Prompt: "Build a KakaoTalk chatbot for cafe order processing. It should support menu browsing, ordering, and payment guidance" Expected result:

• Persona: Bright, friendly cafe staff character, formal tone, active emoji usage
• Conversation design: Menu browsing/ordering/payment guidance/business hours intents, slots (menu name/quantity/options)
• NLU: LLM prompt-based intent classification, menu name entity dictionary construction
• Integration: KakaoTalk chatbot API integration, card-type messages for menu display
• Testing: Order flow happy path, fallback for unavailable menu items, multi-turn ordering

Existing File Reuse Flow

Prompt: "I have this conversation design document; implement the NLU and integrate it with KakaoTalk" + design document attached Expected result:

• Copy existing design document to _workspace/02_conversation_design.md
• Skip persona and designer; deploy nlu-dev + integrator + tester
• Implement NLU based on existing intent/entity catalog

Error Flow

Prompt: "Build me a chatbot" (purpose, channel unclear) Expected result:

• persona proposes a generic persona and requests purpose confirmation from user
• Proceed with remaining pipeline after purpose is confirmed
• Note in report: "Channel undetermined — web widget applied as default"

Agent Extension Skills

Extension skills that enhance agent domain expertise:

Skill	File	Target Agent	Role
intent-taxonomy-builder	.claude/skills/intent-taxonomy-builder/skill.md	nlu-developer, conversation-designer	Intent taxonomy design, entity-slot mapping, training data generation, confusion matrix
conversation-flow-validator	.claude/skills/conversation-flow-validator/skill.md	dialog-tester, conversation-designer	Dialog flow defect detection, Happy/Sad/Edge testing, fallback hierarchy, quality metrics