Creating AI Agents

Beta Feature

AI agents and their capabilities are in active development. Some features like artifacts, tools, and integrations may not work as expected. We're continuously improving reliability.

AI agents are the heart of your coaching app. This guide covers how to create agents that provide valuable, on-brand responses to your members.

Creating an Agent

1. Go to AI Chat and click the Agents menu (🤖 robot icon) in the sidebar
2. Click Create Agent
3. Fill in the agent details
4. Click Create

Quick Edit from Chat

When chatting with an agent, click the Edit button in the chat toolbar to jump straight to that agent's settings. Works for both organization agents (owners/admins) and personal agents (members).

Agent Fields

Field	Description	Example
Name	What members see	"Coach AI"
Avatar	Agent's image	Upload image or use emoji
Description	Brief description	"Your fitness assistant"
Instructions	System prompt for the AI	See below
Model	Which AI model to use	GPT-4o, Claude 3.5 Sonnet
Temperature	Creativity level (0-2)	0.7 default
Knowledge Base	Documents for RAG	Select from your KBs
Persona	Voice and identity	Select or create
Allowed Tiers	Who can access	Select tiers or "All"

Temperature Setting

Temperature controls how creative vs. precise your agent's responses are:

Setting	Value	Best For
Precise	0.0-0.3	Factual answers, technical content
Balanced	0.5-0.7	General coaching (default)
Creative	0.8-1.2	Brainstorming, ideation

Adjust temperature in the agent's Capabilities panel.

AI Models

Choose the right model for your agent:

Provider	Model	Best For	Speed
OpenAI	GPT-4o	Best reasoning, coding	Fast
OpenAI	GPT-4o-mini	Quick responses, cost-effective	Very Fast
Anthropic	Claude 3.5 Sonnet	Detailed analysis, long content	Fast
Anthropic	Claude 3 Opus	Complex reasoning	Medium
Anthropic	Claude 3 Haiku	Quick, economical	Very Fast
xAI	Grok	General purpose	Fast

Personas vs. Instructions

ToolChat separates WHO the agent is from WHAT it does:

Personas	Instructions
Identity and background	Purpose and behavior
Core beliefs (3-5)	Knowledge base rules
Voice characteristics	Response formatting
Banned phrases/styles	Domain-specific rules
< 1,500 characters	< 2,000 characters

Why Separate Them?

When persona and instructions are mixed together, agents become:

• Generic (trying to satisfy too many rules)
• Inconsistent (conflicting directives)
• KB-averse (skip searching when they "feel" knowledgeable)

By keeping them separate, each part can be lean and focused.

Creating a Persona

1. In the agent editor, find the Persona dropdown
2. Click Persona Writer to open the AI assistant
3. Choose your method:
   • Interview — Answer questions about voice and beliefs
   • Transcript — Paste content and extract the voice
4. Review and save the generated persona

Persona Writer

The Persona Writer AI helps you create authentic voice personas:

Interview Mode:

• Asks one question at a time
• Captures name, beliefs, voice, banned phrases
• Generates a concise persona document

Transcript Mode:

• Paste podcast transcripts, interviews, or writing samples
• Extracts speech patterns, filler phrases, beliefs
• Creates a persona that sounds like the real person

Voice Authenticity

The best personas come from real content. If you're creating an agent that sounds like a specific person, provide transcripts or writing samples rather than just describing them.

Writing System Prompts

The system prompt (Instructions) defines WHAT your agent does. It tells the AI:

• Its purpose (1-2 sentences)
• When to search the knowledge base
• Response style guidelines
• What it should never do

Basic Structure

[Identity]
Who is this AI? What's its role?

[Methodology]
What approach does it follow?

[Tone & Style]
How should it communicate?

[Boundaries]
What shouldn't it do?

[Special Instructions]
Any specific behaviors?

Example: Fitness Coach

You are FitCoach AI, an assistant for [Your Name]'s fitness coaching 
program. You specialize in strength training and sustainable nutrition.

## Your Approach
You follow the Progressive Overload methodology:
- Focus on compound movements
- Gradual weight increases
- Adequate recovery
- Flexible nutrition (80/20 rule)

## Communication Style
- Be encouraging and supportive
- Use simple, clear language
- Provide actionable advice
- Celebrate progress, no matter how small

## Boundaries
- Don't diagnose medical conditions
- Don't prescribe specific diets for medical issues
- Always recommend consulting a doctor for health concerns
- Don't make claims about specific results or timelines

## References
When relevant, reference [Your Name]'s programs:
- "Foundations" - Beginner strength program
- "Advanced Gains" - Intermediate program
- "Nutrition Reset" - Nutrition course

Example: Business Coach

You are GrowthBot, the AI assistant for [Business Name]'s 
entrepreneurship coaching program.

## Your Expertise
- Business strategy and planning
- Marketing and customer acquisition
- Sales and revenue growth
- Team building and leadership
- Productivity and time management

## Methodology
You follow the "Lean Startup" approach combined with 
proven business fundamentals:
- Test assumptions quickly
- Focus on customer value
- Data-driven decisions
- Iterate based on feedback

## Response Style
- Be direct and actionable
- Use bullet points for clarity
- Include specific examples when possible
- Ask clarifying questions for vague queries

## Boundaries
- Don't provide specific legal or tax advice
- Don't guarantee specific financial outcomes
- Recommend professional advisors for complex legal/financial matters

Prompt Engineering Tips

Be Specific

❌ "Be helpful" ✅ "When a member asks for a workout, always include sets, reps, and rest periods"

Define Behavior

❌ "Answer questions" ✅ "If you don't know something, say 'I'm not sure about that specific topic' rather than making something up"

Set Boundaries

❌ (nothing) ✅ "Never provide medical advice. If asked about injuries or health conditions, recommend consulting a healthcare professional."

Include Examples

When a member says they're struggling with motivation, respond with:
1. Acknowledge their feelings
2. Ask what specific challenge they're facing
3. Offer 2-3 practical strategies
4. End with encouragement

Artifacts

Enable artifact generation to let AI create:

• Documents — Plans, guides, outlines
• Code — Scripts, snippets, examples
• HTML — Landing pages, forms
• React — Interactive components, calculators

Enabling Artifacts

1. Edit your agent
2. Toggle Artifacts Enabled
3. Optionally add Artifact Instructions for specific guidance

Artifact Instructions Example

When creating workout plans, generate them as formatted documents.
When creating calculators (BMI, macros, etc.), use React components.
Always make artifacts visually appealing and easy to read.

Conversation Starters

Add suggested prompts that show when members start a chat:

Starter	Purpose
"Help me create a workout plan"	Direct action
"What should I focus on first?"	Orientation
"I'm struggling with motivation"	Support
"Explain progressive overload"	Education

Good starters:

• Address common needs
• Reduce blank page syndrome
• Guide members to valuable interactions

Tier Access

Control which tiers can access each agent:

• All Tiers — Everyone with AI Chat access
• Specific Tiers — Only selected tier members

Use this to:

• Offer premium agents to higher tiers
• Create tier-specific specialists
• Add value to upgrades

Multiple Agents

Consider creating multiple agents for different purposes:

Specialist Agents

Agent	Purpose
General Coach	Broad questions, orientation
Workout Planner	Exercise programming
Nutrition Guide	Diet and meal planning
Mindset Coach	Motivation and psychology

Benefits of Specialists

• More focused responses
• Better at specific tasks
• Clearer purpose for members
• Easier to refine prompts

Testing Your Agent

Before launching:

1. Test common questions — Ask typical member questions
2. Test edge cases — Try questions outside your domain
3. Test boundaries — Verify it handles restricted topics properly
4. Test tone — Ensure responses match your brand voice
5. Test artifacts — If enabled, verify output quality

Sample Test Questions

• "How do I get started?"
• "I'm feeling unmotivated"
• "What should I eat today?"
• "Create a weekly plan for me"
• "Can you diagnose why I'm tired?"

Refining Agents

Agents improve over time:

1. Gather feedback — Ask members what's working/not working
2. Review conversations — Look for patterns in suboptimal answers
3. Update prompts — Add new instructions based on findings
4. Test changes — Verify improvements before going live
5. Update knowledge base — Add missing information

Version History

Every time you update an agent's system instructions, the previous version is saved:

To view version history:

1. Edit your agent
2. Click Version History next to the Instructions field
3. See all previous versions with timestamps

Using version history:

• Compare — See what changed between versions
• Restore — Roll back to a previous version if needed
• Track improvements — Monitor how your prompts evolve

Best Practice

Before making major changes to system instructions, note what the current version does well so you don't lose effective patterns.

Fine Tune (Quick Improvements)

See a response that's not quite right? Fix it instantly without leaving the chat.

How Fine Tune Works

1. Find a response you want to improve
2. Click the Fine Tune button on that message
3. A drawer opens with the Fine Tune Assistant
4. Describe what should change
5. Review the proposed patch (shown as a diff)
6. Click Apply to update the agent

What Fine Tune Does

• Surgical patches — Adds one specific rule without restructuring
• Preserves structure — Matches your existing formatting
• Shows changes — Preview diff before applying
• Saves version — Previous instructions are preserved in history

When to Use Fine Tune vs. AI Writer

Fine Tune	AI Writer
One specific fix	Full rewrite or restructure
"Never say X" rules	New agent from scratch
Quick behavioral tweak	Major instruction changes
From within the chat	From the agent editor

Example Fine Tune Workflow

You see: Agent gives generic advice instead of using KB content

You click: Fine Tune button

You type: "When asked about sales scripts, always search the KB first"

Assistant shows:

+ - When asked about scripts, frameworks, or techniques: ALWAYS search knowledge base first

You click: Apply

Done! The agent now has that rule baked in.

AI Writer

Need to create new instructions or do a major rewrite? Use the AI Writer.

Accessing AI Writer

1. Open the agent editor (create new or edit existing)
2. Click AI Writer button
3. Chat with the AI Writer assistant

What AI Writer Does

• Interviews you — Asks about purpose, audience, rules
• Generates lean instructions — Under 2,000 characters
• Emphasizes KB usage — Forces agents to search before answering
• Follows best practices — No framework definitions inline

AI Writer Best Practices

The AI Writer follows these principles:

DO include:

• Purpose (1-2 sentences)
• KB trigger rules (specific phrases that force search)
• Response style guidelines
• "Never do" rules

DON'T include:

• Framework definitions (put in KB)
• File listings (AI Search handles this)
• Lengthy explanations
• Duplicate rules

Keep Instructions Lean

Instructions over 2,000 characters often cause agents to skip KB search and give generic responses. Let your Knowledge Base do the heavy lifting.

Common Issues & Fixes

Issue	Solution
Too generic	Add more specific examples to prompt
Too verbose	Add "Be concise" or "Keep responses under X words"
Missing info	Upload more to knowledge base
Wrong tone	Add explicit tone instructions or use a Persona
Off-topic responses	Add clearer boundaries
Poor artifacts	Add artifact instructions

Agent Order

Drag and drop agents in the list to reorder them. The order affects how they appear to members in the agent selector.

Default Agent

Set one agent as the default. This agent is selected automatically when members open AI Chat.