tannhatcms/jarvis/ams-ai-chatbot-builder

AI Chatbot Builder

A good chatbot is a designed conversation, not a model with a text box. This skill covers the full build: defining the job, designing the dialogue, grounding answers, handling failure gracefully, and shipping.

Overview

Start from the job: what questions should this bot answer, and what actions should it take? Design the conversation — greeting, disambiguation, answer, follow-up — before writing prompts. Ground answers in your knowledge base so the bot says true things. Plan the failure paths: confusion, out-of-scope requests, angry users, and the handoff to humans. Then ship with analytics, because the first version is a hypothesis about what users will ask.

When to use

• Customer support, sales assistance, or internal helpdesk automation.
• FAQ-style bots where answers exist in documentation.
• Lead qualification or intake bots with structured data capture.
• Replacing or upgrading a rules-based bot with an LLM-powered one.

Core concepts

• Conversation design: the happy path plus the top failure paths, scripted as flows. Greeting

→ intent → resolve → confirm → close. Design the unhappy paths with equal care.

• Persona: the bot's voice — tone, formality, name, boundaries. Consistent persona builds

trust; document it so every prompt matches.

• Grounding: answers built from retrieved sources (docs, KB articles) with citations. The

difference between a helpful bot and a confident liar.

• Intent handling: recognize what the user wants — including multi-intent and ambiguous

messages — and disambiguate explicitly rather than guessing.

• Escalation: the human handoff — when (confidence low, sensitive topic, user asks), and how

(with full context transferred, not a cold restart).

• Analytics: track intents, resolution rate, escalation rate, user satisfaction, and the

"unknown" bucket. The unknown bucket is your roadmap.

Practical workflow

1. Define scope: the top 20 questions/intents from real user data (tickets, logs). If you don't

have data, interview the team that talks to users.

1. Build the knowledge base: clean, current source documents; chunk and index for retrieval.
2. Design conversation flows for the top intents plus failure paths (confused, out-of-scope,

frustrated).

1. Implement with grounding: retrieve → generate with citations → confidence check → answer

or escalate.

1. Test with real user phrasings — messy, typo'd, ambiguous — not your clean test cases.
2. Launch to a slice of traffic; review the unknown bucket weekly; expand scope based on data.

Bot spec template:
SCOPE:    <top intents — from real data>
PERSONA:  <name, tone, formality, boundaries>
SOURCES:  <knowledge base docs + freshness owner>
FLOWS:    <happy paths + confusion + out-of-scope + escalation>
ESCALATE: <triggers + context handoff format>
METRICS:  <resolution %, escalation %, CSAT, unknown-bucket size>

Common pitfalls

• Scope creep at launch: trying to answer everything on day one. Launch narrow, expand from the

unknown bucket.

• Ungrounded answers: the bot improvising from model memory. Ground everything; abstain when

sources are silent.

• No escalation path: trapping frustrated users in a bot loop. Always offer a human, early.
• Testing with clean inputs: real users are messy. Test with typos, slang, multi-intent

messages.

• Ignoring the unknown bucket: the list of unhandled questions is the highest-value analytics

you have. Review it weekly.

• Set and forget: knowledge goes stale, intents drift. Assign a KB owner and a review cadence.