Structuring high-quality instructions is key to ensuring your AI responds accurately, whether used for AI Agents or Recommendations. This best practices guide will help you get started.

💡

AvoAI fees are deducted in real-time from your Account Balance. Realtime billing is available in Analytics and on the Billing page.

Use Recipes to Get Started

Using a recipe ensures you follow best practices from the beginning.

Start with one of the pre-set recipes available on the Instruct Agent section. These are templates that auto-populate common structures that you can customize to match your brand voice, workflows, and logic rules.

Click the Recipe button to open available templates

Make a selection

Customize

Save and enable the instruction

You can continue to edit, enable or disable all instructions after the first pass.

Follow These Best Practices

To ensure reliable and maintainable AI behavior, always follow the rules below:

Start your Instruction with any rules first.

Use explicit, plain logic. Avoid nesting. Be as clear as possible.

Write conditions in plain language, for example: "If the customer mentions billing, send them the billing FAQ." No capital letters or special syntax needed.

Use double quotes (") for exact response text to clearly delineate the bot’s reply.

Consolidate logic into one section. Do not repeat rules or logic.

Use markdown headers and lists to structure, e.g. 1, 2, 3, and organize the instruction clearly.

Give the reason. A short "because" helps the agent handle cases you didn't list. Compare "Don't quote prices" with "Don't quote prices, because they vary by location and our team confirms them."

Decide when to script and when to guide. Script exact wording for links, pricing, and compliance language. For everything else, describe the goal and let the agent phrase it.

Say what to do, not only what to avoid. "Reply in short, plain sentences" works better than "Don't write long replies."

Always include a fallback. Tell the agent what to do when a message matches none of your rules, such as asking one clarifying question or handing off to a teammate.

Be specific about links. Instead of "Don't reference internal links," write "Only share links that appear in these instructions."

Tell the agent when to apply a tag. The agent only applies tags that are on its permitted list, and only when your instructions say when. Tags are case-sensitive, so match the spelling exactly.

Keep instructions current. Update them when your hours, prices, or policies change.

Instructions vs. Settings

Your agent's behavior lives in two places. Keeping them separate keeps your instructions short and predictable.

Instructions control business logic and wording: the agent's role, the order it works through steps, what to collect, when to hand off and what to say, and its guardrails.

Configure Agent settings control platform behavior: agent name, tone preset, response delay, office hours, conversation memory, transfer to a human, whether the agent skips a final message on escalation, and Insights.

Rule of thumb: logic and wording go in Instructions, while routing and platform behavior go in settings. For escalation, see AI Agent Escalation and Assignment.

Add Templates, Tags or Custom Fields into Instructions

When creating an instruction, use the icons to quickly reference and add your existing templates, tags and custom fields into your instructions.

Draft Instructions with the Avochato MCP Connector

You don't have to write instructions from a blank page. With the Avochato MCP connector, an AI assistant such as Claude can read your real conversation history. Combine that with a description of the agent you want and this guide, and you can draft instructions grounded in how your team already talks to customers.

Connect the Avochato MCP connector to your AI assistant.

Pull in real conversations. Ask the assistant to review recent conversations for the topic you want to automate: the most common questions, how your best replies are worded, and which situations your team escalates.

Describe the agent you want. Say what its role is, what success looks like, what it should collect, when it should hand off to a person, and what it must never do.

Share this article as the guide. Ask the assistant to write instructions in the structure described here (rules first, if/then logic, exact quoted replies, a fallback), reusing the wording of your strongest real replies. For an intake or qualification flow, also share How to Build a Triage Agent.

Review and edit. Treat the output as a first draft. Check every link, price, and policy for accuracy.

Test before going live. Paste the draft into a new instruction and try it in the Test Agent Sandbox.

Tips

Conversation history can contain customers' personal information. Use it to learn patterns and wording, and don't paste names, phone numbers, or account details into your instructions.

Start with one topic at a time, such as billing questions or appointment requests, and create one instruction for each.

Ask the assistant to point out conversations that don't fit any of your rules. Those are usually the gaps your fallback should cover.

Example Instruction Structure

Here’s how to format an effective AI instruction:

Role

You are a customer support agent at Acme Corp.

Objective

Your goal is to help users troubleshoot account issues and refer them to support documentation when needed.

Rules

If the user mentions login problems, reply with: "I'm sorry you're having trouble logging in. Please try resetting your password here: https://support.acme.com/reset-password"

If the user mentions billing, direct them to: "Our billing FAQ is available at https://support.acme.com/billing. Let me know if you’d like help with a specific charge."

Only share links that appear in these instructions. Don't reference internal tools or links, because customers can't open them.

Respond in a professional and friendly tone.

Fallback

If the message doesn't match any rule above, ask one clarifying question.

If the reply still doesn't match, hand off to a teammate.

Instruction Fields Explained

1. Instruction Name

Give your instruction a clear, internal-facing name (e.g. “Billing Troubleshooting” or “Sales Chat Intro”).

This helps you locate and manage instructions later. The name isn't used by the AI Agent in its replies. Only the instruction content is.

2. Content

This is where the training happens. Add detailed instructions for your bot’s tone, purpose, workflows, and rules. There is no character limit, so more context is better.

Use:

Structured markdown (#, -, etc.) to organize sections

Explicit instructions and logic

Do include clear response scripts and conditional logic.

Do not leave rules vague or embed logic in prose paragraphs.

Using IFTTT-Style Logic

Use if this, then that (IFTTT) statements to define exactly what the bot should say and when. This ensures predictable responses and makes your instruction easier to update.

Example:

If the user asks about refunds, respond with: "You can request a refund by filling out this form: https://support.acme.com/refunds"

Copy-and-Fill Template

Paste this into a new instruction, replace the bracketed parts, and delete any section you don't need.

# Role
You are [role] at [Company]. Your goal is to [what success looks like].

# Tone
[Describe the voice, for example friendly, concise, plain language. Add length guidance if useful.]

# Rules
- If [what the contact says or asks], reply with: "[exact wording]"
- If [another situation], [what the agent should do], because [reason].
- Only share links that appear in these instructions.
- Don't [off-limits topic], because [reason].

# Fallback
- If the message doesn't match any rule above, [ask one clarifying question / hand off to a teammate].

Example Instruction Sets

Each example works as its own instruction, so you can enable or disable them independently. For a complete multi-step intake flow, see How to Build a Triage Agent.

1. Role and tone

You are a customer support agent at Acme Corp. Your goal is to resolve common account questions quickly and point people to the right resource. Be professional and friendly. Use short sentences and avoid jargon.

Why it's built this way: one short paragraph that sets the role, goal, and tone. Everything else builds on it.

2. Hours and location

If the contact asks when we are open, reply with: "We're open Monday to Friday, 9am to 5pm Eastern."
If the contact messages outside those hours, let them know a teammate will follow up the next business day.
If the contact's first name is available, use it in the greeting. If it is not, skip the name. Never guess a name.

Why: exact wording for the facts, plus a rule for a missing field. Using contact fields requires the agent's contact data access setting to be on.

3. Lead qualification

If the contact is interested in buying, ask these one at a time: what they are looking for, their timeline, and their rough budget.
Don't ask all three in one message, because long messages get fewer replies.
When all three are answered, apply the tag "Qualified Lead" and hand off to a teammate.

Why: one question at a time keeps replies short, and the tag fires at a clear moment.

4. Applying tags

If the contact says they want to cancel, apply the tag "Cancellation Risk".
If the contact asks for a refund, apply the tag "Refund Request".
Only use these exact tags. Never create new ones.

Why: tags are case-sensitive and the agent only applies tags it's been told about, so exact names matter.

5. Escalating to a person

If the contact asks for a human, sounds frustrated, or mentions a legal or safety issue, don't try to resolve it yourself. Send exactly: "Thanks for letting us know. I'm bringing in a teammate to help." Then hand off.

Why: these situations need a person's judgment. Sending the message first means the contact never feels ignored.

6. Boundaries and compliance

Don't give medical, legal, or financial advice, because the answer could be wrong for their situation. If asked, say a teammate can help and hand off.
Don't quote prices or promise discounts, because pricing varies and our team confirms it.

Why: each guardrail has a reason, so the agent handles near-miss questions correctly.

7. Fallback

If the message doesn't match anything above, ask one clarifying question.
If the reply still doesn't fit, hand off to a teammate. Never guess.

Why: without a fallback the agent has to improvise.

Test Before You Go Live

Use the Test Agent Sandbox to simulate conversations without affecting your live inbox.

Assign the agent to yourself and start a test conversation to confirm the full flow, including tags and handoffs.

Try messy inputs: typos, off-topic questions, two questions in one message, and messages that match two rules at once.

Confirm routing and auto-assignment send escalated conversations to the right team.

Start with a longer response delay (for example 60 seconds) so your team can preview and cancel messages while you refine.

Review early conversations after launch. If the agent keeps getting something wrong, clarify or reposition that rule rather than repeating it.

The agent only joins conversations it's assigned to, either manually or through auto-assignment and skill-based routing.

Managing Instructions

Enable / Disable

Enable an instruction once it’s ready—your AI Agent will start using it right away.

Disable instructions any time to pause usage without deleting them.

Save or Update

Click Save when creating a new instruction.

Click Update after editing an existing one (hover and edit inline).

Organize with Multiple Instructions

You can create multiple instructions to split different functions, such as:

Sales responses

Live chat troubleshooting

Onboarding flow

This lets you disable specific areas for updates without disrupting others.

Summary

High-quality prompts are:

Structured with clear rules and markdown

Written with explicit logic in plain language

Backed by a reason, a fallback, and copy-paste-ready replies

Drafted from real conversations, then tested in the sandbox

Modular and easy to manage

Questions? Reach out to Avochato Support for further guidance in training your Avochato AI Agent to be consistent, effective, and aligned with your business goals.

Best Practices for Instructions