sf-ai-agentforce-conversationdesign_skill

Overview

This skill brings conversation design methodology to Salesforce Agentforce. It generates persona documents, topic architectures, instruction sets, utterance libraries, escalation matrices, and guardrail configurations. It also validates existing agents against a 120-point quality scorecard and produces prioritized improvement plans.

How this skill works

The skill analyzes agent-level artifacts and runtime design (persona, topics, instructions, utterances, escalations, and guardrails) and scores each area across eight categories totaling 120 points. It generates concrete deliverables: persona templates, topic maps, three-level instruction sets, utterance libraries, escalation matrices, and guardrail configs. For existing agents it runs the quality assessment, highlights gaps by criterion, and outputs a prioritized improvement plan with actionable fixes and example text.

When to use it

• Design a new Agentforce agent from scratch with consistent persona and scope
• Audit and score an existing agent to identify design regressions or safety risks
• Create production-ready instruction sets and utterance libraries for reliable classification
• Define escalation and handoff behavior with clear routing and context handoffs
• Establish guardrails, PII handling, and deterministic safety rules before deployment

Best practices

• Start bottom-up: list actions, then group into topics with ≤5 actions each
• Write an explicit persona in agent-level instructions (200–500 words) to set tone and limits
• Use the three-level instruction model: agent, topic, action with recommended lengths
• Keep topic count focused (≤10) and ensure classification descriptions are semantically distinct
• Provide ≥5 natural phrasings per primary intent and include edge/adversarial utterances
• Separate business logic (Flows/Apex) from guidance; keep determinism in code, not in instructions

Example use cases

• Generate a service-agent persona, welcome and fallback messages, and full instruction set for a SaaS support bot
• Assess a retail agent’s topic architecture and produce a 120-point score with itemized remediation tasks
• Create an utterance library including synonyms, misspellings, and out-of-scope examples for training and testing
• Build an escalation matrix that maps triggers to Omni-Channel routing, priority levels, and handoff messaging
• Define guardrail hierarchy and PII handling rules to prevent hallucination and ensure deterministic safety