prompt-review

Prompt Review Skill

Purpose

Review an AI agent prompt before it is tested, deployed, or edited in production. Focus on whether the prompt is reliable, specific, safe, aligned with available tools/actions, and suitable for real customer conversations.

Especially useful for Synthflow voice agents, Single-Prompt agents, Flow Designer node prompts, transfer flows, escalation logic, support/sales/qualification/booking agents, and any prompt that controls call behavior.

Getting the material over MCP

When the prompt lives in a Synthflow workspace rather than in the conversation, pull it with the Synthflow MCP tools: get_agent for the live prompt and settings, get_agent_actions / get_action for the action descriptions the prompt must agree with, and list_agent_versions / get_agent_version_diff when comparing an edit for regression risk. To check how a Synthflow feature actually works, search the official docs with search_docs, or with the searchDocs tool on the synthflow-docs server if the workspace tools aren't connected yet.

Core Review Philosophy

A good prompt is understandable by a human operator with no hidden context.

The "another human" test (OpenAI's recommended check): read the prompt aloud to a teammate. If they can't restate what the agent should do in plain English, the LLM won't reliably do it either.

Write so an 18-year-old with zero context could execute it the same way twice. Prefer concrete behavioral instructions over vague personality instructions.

• Bad: Be helpful and professional.
• Better: Use one or two sentences per turn. Ask one question at a time. If the caller asks about pricing, give the approved price range and ask whether they want to book a consultation.

Replace vague verbs with specific, testable behavior:

❌ Vague	✅ Specific
"Handle the caller professionally"	"One sentence per turn, mirror their tone, no slang"
"Help them find the right battery"	"Ask year/make/model/trim, then present 2 in-stock SKUs with price + warranty"
"Verify the customer"	"Collect name, phone, order ref. Read each back. Ask 'is that correct, yes or no?'"
"Be helpful"	(delete — not actionable)