AI Prompting 101 for Sparkforge Conversation AI

The bot is only as good as the prompt. A vague prompt produces a vague bot. A specific prompt with structure, examples, and guardrails turns Conversation AI into a useful sales or support agent that stays on script.

The three elements of a prompt

Role — who the bot is

Set the persona, voice, and context. The bot's tone, language style, and self-awareness all hinge on this. A clear role prevents the bot from sounding generic or breaking character.

Task — what to do

State the objective and the steps. Vague tasks lead to hallucinations and off-script answers. Spell out the script flow.

Guidelines — hard rules

Guardrails the bot must respect: reply length limits, when to escalate, what not to say. These keep the bot inside its lane.

Role: example contrast

Worse: “You are a salesperson for {{ai.business_name}}, assist customers and help them book appointments. Act like a human. You are talking to {{contact.name}}.”

Better:

- You work for {{ai.business_name}}. Your goal is to assist customers, qualify them based on the answers they give to the script questions, and — if qualified — help them book appointments.
- Always maintain the persona of a human assistant. Don't disclose, suggest, or hint that you're an AI unless asked. Use phrases like “Great,” “Got it,” or “Understood.”
- Acknowledge the user's input by repeating it back once at the start, then use varied affirmations afterward.
- Keep tone professional but casual. Use periods, not exclamation points.
- You are talking to {{contact.name}}. Use their first name only at the start and end of the conversation.

Task: example contrast

Worse: “You are required to qualify customers by asking them questions then asking them to book an appointment.”

Better:

Script Flow:
- Open with “What kind of marketing solution are you looking for?” If not interested, ask politely for their email and offer to stay in touch.
- If positive, ask “What telephony provider do you use right now?”
- Once they've answered all script questions positively, propose scheduling a call.
- Before booking, ask for their email if you don't have it.
- After booking, thank them and end on a positive note.

Guidelines: example contrast

Worse: “Reply in a concise manner.”

Better:

- Keep answers short, direct, and within a 20-word limit.
- Always wait for the user's response before asking the next question.
- Only book an appointment if the user responds positively to all script questions.

Best practices

Repetition

Reinforce critical instructions in more than one section. If the bot must follow the script, mention the script in Role and in Guidelines. The model weights repeated instructions higher.

Examples

Show the bot what to say (and what not to). Concrete examples beat abstract instructions.

EXAMPLES OF WHAT TO SAY AND WHAT NOT TO SAY:
- Avoid: “I didn't understand your response.”
- Use: “Wait, what did you say? Sorry, could you repeat that?”
- Avoid: “I apologize for the confusion.”
- Use: “Sorry if that didn't make sense.”

Iterative improvement

Treat the prompt as a living document. Use the Bot Trial tool inside Conversation AI to test, tweak, and ship. Run the same conversation through the bot multiple times to surface inconsistency. Small wording changes can shift behavior — confirm with multiple test runs before declaring a win.

Advanced techniques

Static information

Drop business context, offers, or services directly into the prompt so the bot has them on hand. Keep this block under 100–200 words; longer context dilutes the rest of the prompt and the bot's adherence drops.

Delimiters

Use special characters to mark sections so the model treats them as discrete blocks rather than free text:

• Emphasis: #, >, <
• Demarcation: """, ''', <tag> </tag>

Example wrapping an offer block:

# Offer:
<offer>
[Your offer copy here.]
</offer>

Conversation context

Tell the bot why someone is reaching out. Instead of “You are a customer support chatbot,” try: “A client is contacting us because something went wrong with their order. Act as a friendly agent collecting a clear description of what happened.”

Writing styles and tones

Pick deliberately:

• Styles: Analytical, Conversational, Creative, Descriptive, Informative, Instructive, Persuasive, Technical.
• Tones: Authoritative, Confident, Empathetic, Formal, Friendly, Playful, Serious, Warm.

Mismatched style and tone confuse the bot. A “technical” style with a “playful” tone often produces awkward output — stick to combinations that match the audience.