Prompt – Instructions for your Agent

What is the Prompt?

The prompt is the core set of instructions that your Agent follows every time it works on a task. Think of it like the instructions you’d provide to an employee about how they should work. For example: their responsibilities, how they should communicate, what rules they must follow, scope of their tasks, etc.

Agent Profile

Agent Name

Give your Agent a clear, memorable name that reflects its purpose or role.

• Good: “Invoice Processor Agent” – immediately implies what it does.

• Avoid: “Bot123” or “HelperAgent” which are vague.

Agent Summary

Write a short, descriptive summary of what the Agent does and its area of expertise. This helps the LLM behind the Agent to understand its role, and it helps team members know what to expect. Example: “This Agent reviews incoming invoices, extracts line items, matches to purchase orders, flags discrepancies and posts summary to the accounting system.”

Model Selection

Choose which LLM your Agent should use (or let Arahi AI pick automatically). Key considerations:

• Cost vs performance trade-off.

• Use higher-performing models for tasks requiring reasoning, domain knowledge or sensitive data.

• Use lighter models for simpler tasks or high-volume runs.

Prompt Instructions

Invent a Prompt (Quick Start)

If you want to get started fast, use our “Invent” feature: just describe what you need, and Arahi AI will generate a structured prompt for you automatically.

Write Your Own Prompt (Full Control)

If you prefer full control, craft your prompt manually. Key parts to include:

• Role definition: “You are _AgentName, a specialist in domain.”

• Responsibilities: What tasks you will perform, what you will not.

• Input format: What kind of input will you receive? e.g., “client name, invoice PDF, PO number”.

• Output format: What you must deliver: JSON, bullet list, table, Slack message.

• Boundaries and Escalation: When to stop, when to ask for human help.

• Tone & Style: Professional? Friendly? Factual? Creative?

• Examples: Provide few-shot examples if possible (helpful for LLMs).

• Constraints: e.g., “Do not expose PII”, “Keep answer under 300 words”.

Writing Tips

• Be as clear and specific as possible: vague instructions lead to ambiguous responses.

• Use active voice and shorter sentences.

• Set expectations: show what good output looks like.

• Anticipate edge cases: “If you cannot determine the PO number, escalate to human”.

• Use variables/placeholders if you want dynamic content (e.g., {{client_name}}).

• Test with sample inputs and refine accordingly.

Example Prompt Template

“You are InvoiceProcessor-Agent. Your job is to receive an invoice PDF, extract the supplier name, invoice date, line-items (description, quantity, price), match against purchase orders in our ERP. If any line-item does not match a PO, flag it and escalate to the finance team. Output a JSON object with keys: supplier, invoice_date, line_items, matched_PO, mismatch_flag. Use friendly, professional tone. If you cannot find a matching PO, add mismatch_flag: true and send a Slack alert.”

Testing & Refinement

• Run test inputs through the Agent to see how it performs.

• Check for clarity, accuracy, tone, edge-cases.

• If output is incorrect or inconsistent, revise your prompt (or add guiding examples).

• Monitor over time: Does the Agent drift? Is it using the wrong tone? Does it mis-interpret inputs? Plan for periodic prompt refinement.