Step 6 of 10

Compose constrained reply drafts

Generate short, useful reply drafts with consistent tone, one per intent category, using separate prompt templates that match the routing policy.

Why this matters

A single generic prompt cannot write a good reply to both 'I have a general question about your pricing page' and 'I need urgent technical support, my account is locked.' Each intent category requires a different tone, different expected information, and different call to action. Separate prompt templates enforce this.

Build instructions

Build the inquiry reply template

This template handles general questions about your service, product, or availability.

Step 1
ROLE: You are a professional customer experience specialist. Write a first reply email body only, no subject line, no greeting beyond the first line.
Step 2
CONTEXT: Sender name: [Full Name]. Their message: [Normalized body]. Intent: General inquiry.
Step 3
OUTPUT STRUCTURE: Line 1: Personalized greeting using their name. Paragraph 1: One sentence acknowledging their specific question topic. Paragraph 2: One sentence confirming the team will follow up with details. Paragraph 3: One action, either a resource link or a specific next step.
Step 4
FORBIDDEN: Do not fabricate prices, availability, or timelines. Do not promise anything not stated in their message. Do not exceed 100 words. Do not use filler phrases like 'Great question!' or 'Absolutely!'

Add a Paths step to route to the correct template

Step 1
After the safety guardrails Path A, add another Paths step inside the auto-send path. Create one path per intent category that can be auto-sent.
Step 2
Each path condition: Intent equals 'inquiry' → use inquiry template. Intent equals 'follow-up' → use follow-up template.
Step 3
Each path contains one OpenAI action step with its specific prompt. All share the same Max Tokens setting (200-250) for consistent reply length.

Test each template separately

Step 1
For each intent-specific template, run at least two test emails through its path. Evaluate: Does the reply acknowledge the specific intent? Is it under the word limit? Does it avoid all forbidden content?
Step 2
Compare the output across different input messages for the same intent. The structure should be consistent even though the content varies.

Common mistakes

Using one generic prompt for all intent categories. 'General question' and 'Follow-up' require completely different tones and information. A generic prompt produces generic output that feels unhelpful.
Not setting Max Tokens per template. Without a token limit, the model may write 400-word essays for simple inquiries. Set 200-250 tokens for short reply drafts.

Pro tips

Write the ideal human reply for two emails in each category before writing the AI prompt. Then reverse-engineer what instructions would produce that output. This is the fastest way to write effective prompts.

Before you continue

Each active template has been tested with at least two real emails. All outputs are under 150 words, address the specific intent, and contain no fabricated content. You could send any of the test outputs without editing.

Step result

Each intent category has a dedicated, tested prompt that produces appropriate, constrained replies for that specific type of email.