Overview
Lyra AI Studio is the admin workspace for creating, configuring, and maintaining Lyra coaches. From here, admins can author new coaches, curate their knowledge, preview behaviour, control access, and manage ongoing updates.
Access to Lyra AI Studio is permission-based. Ensure your account has the relevant admin permission before proceeding. See the Access and permissions section below.
On this page: Access and permissions | Create a new Lyra coach | Configuring a Lyra coach | Knowledge | Access | Settings | Personality | Policy | Options | LLM (model selection) | Preview | Troubleshooting | FAQs
Access and permissions
Who can access Lyra AI Studio
-
Full site admins
-
Members of an admin group with the Lyra/AI Coach management permission enabled
Grant or revoke access centrally via Admin groups. Once permission is enabled, eligible users see Lyra AI Studio in the Admin Panel and can create and edit coaches.
How to grant access
-
Open the Admin Panel and select Admin groups.
-
Go to Group permissions and pick the target admin group.
-
Enable the AI Coach management (Lyra AI Studio) permission and click Save.
All users in that group will now see Lyra AI Studio and can manage coaches.
Create a new Lyra coach
-
In Lyra AI Studio, click Create coach.
-
Complete the required fields:
-
Name — A clear, user-facing title (e.g. "Onboarding Coach")
-
Description — A short summary users will see when browsing available coaches
-
Purpose — What the coach is for and how it should behave
-
Icebreaker — Instructions for the opening message users see
-
-
Click Save. The new coach appears in the coaches list with a Live status indicator.
Once created, you can fully configure the coach using the editor. Select the coach from the list and click Edit to access the three main configuration areas: Knowledge, Access, and Configure.
Configuring a Lyra coach
The coach editor is organised into three top-level tabs. Each tab controls a different aspect of how the coach behaves and who can use it.
|
Tab |
What it controls |
|---|---|
|
Knowledge |
Which communities the coach draws content from to answer questions |
|
Access |
Which communities' members can see and interact with the coach |
|
Configure |
The coach's identity, personality, policies, chat options, AI model, and live preview — split across six sub-tabs: Settings, Personality, Policy, Options, LLM, and Preview |
Knowledge
The Knowledge tab controls which Fuse communities the coach uses as its knowledge base. The coach will have access to the contents in the selected communities and base its answers on those contents.
How to configure
-
Open the Knowledge tab in the coach editor.
-
Use the search field to find communities, or browse the list.
-
Select one or more communities by ticking the checkbox next to each one. Selected communities appear as pink tags below the list.
-
Click Save.
Important: Choose communities deliberately. A smaller, well-curated set of knowledge sources will outperform a broad, unfocused selection. Only include communities that contain content relevant to this coach's purpose.
For detailed guidance on supported file types, how semantic search works, and best practices for knowledge base curation, see Lyra knowledge base.
Access
The Access tab controls who can see and interact with the coach. The coach is accessible to any user who has access to the selected communities.
How to configure
-
Open the Access tab in the coach editor.
-
Search for or browse communities in the list.
-
Select the communities whose members should have access to this coach. Selected communities appear as pink tags below the list.
-
Click Save.
Note: The Knowledge and Access community selections are independent. You can give a community access to a coach without that community being a knowledge source, and vice versa. This means you can curate knowledge from one set of communities while making the coach available to a different audience.
Settings
The Settings sub-tab (under Configure) defines the coach's core identity.
|
Field |
Description |
|---|---|
|
Agent Name |
The display name shown to users (e.g. "Product Intelligence Coach", "Onboarding Coach") |
|
Agent Description |
A short summary of what the coach does — visible to users when browsing available coaches (max 400 characters) |
|
Agent UID |
A unique identifier for the coach, auto-generated. Use this when referencing the coach in integrations or support requests. |
|
Agent Purpose |
The core instruction that shapes how the coach responds — its role, scope, tone, and approach. Think of this as a job description for the coach. |
Tip: The Purpose field is the single most important factor in coach quality. Be specific about what the coach should do, how it should behave, what sources to draw on, and when to escalate. See Lyra coach settings for detailed guidance and examples.
Personality
The Personality sub-tab controls the tone, style, and formatting of the coach's responses.
Agent personality
Choose from four predefined personality options, or define a custom one:
|
Option |
Best for |
|---|---|
|
Friendly & Empathetic |
Learning, coaching, and onboarding use cases. Warm, understanding, and supportive. |
|
Professional & Knowledgeable |
Professional writing or important responses. Confident and authoritative. |
|
Insightful & Analytical |
Problem solving, ideation, and research. Detail-oriented, analytical, and data-driven. |
|
Strict & Precise |
Compliance, guidelines, and procedures. Firm, concise, and meticulous. |
|
Custom Personality |
Define your own personality in a free-text field when the presets don't match your needs. |
Output formatting
Controls how the coach structures its responses:
|
Option |
Style |
|---|---|
|
Conversational |
Short paragraphs and emojis. Relaxed and informal. |
|
Structured |
Structured bullet points with a conversational tone. Occasionally uses emojis. |
|
Storytelling |
Analogies, vivid descriptions, and anecdotes. For creative or narrative use cases. |
|
Technical |
Short, simple, and concise. For compliance, policy, or technical documentation. |
|
Custom Formatting |
Define your own formatting instructions in a free-text field. |
Icebreaker message instructions
The icebreaker defines how the coach greets users when they open a new conversation. Write instructions rather than a fixed script — this lets the coach personalise the greeting while staying consistent with its personality.
See Lyra coach settings for detailed icebreaker guidance and examples.
Policy
The Policy sub-tab controls guardrails and behavioural rules for the coach. Use these settings to define scope, manage edge cases, and ensure safe, on-topic responses.
Core policy controls
|
Setting |
What it does |
|---|---|
|
Focus and Scope |
Keeps the coach within its defined content domain. When enabled, you define the scope in a text field — e.g. "Fuse Universal product strategy and prioritisation — specifically: the live Jira backlog, PM frameworks, and client feedback." |
|
Theme-specific Allowance |
Allows the coach to answer questions on a specific theme using general knowledge, even when the answer isn't in the knowledge base. Useful for contextualising data. |
|
Knowledge Limits and Apologies |
When enabled, the coach will be transparent about not having information rather than guessing. Builds trust. |
|
Non-Context Segment Message |
A custom message injected when no matching content is found in the knowledge base. Helps prevent hallucination — e.g. "I don't have matching data for that. Try rephrasing, or check with the Product team directly." |
|
Clarification and User Understanding |
Ensures the coach checks that users understand the information provided, with mechanisms for rephrasing. |
|
Recommendations and Examples |
Allows the coach to provide multiple recommendations or examples where relevant. |
|
Neutrality |
Ensures neutrality when referring to individuals or authors, avoiding gender-specific references. |
|
Professional Conduct and Bias |
Maintains professionalism and ensures responses are free from bias and unprofessional language. |
|
Avoid Legal and Financial Advice |
Prevents the coach from providing legal or financial advice. Encourages consulting qualified professionals. |
|
Deflect Safeguarding and Sensitive Issues |
Ensures the coach does not engage in sensitive topics or safeguarding issues. Instead, it directs users to contact the provided email for support. |
|
Feedback Encouragement |
Encourages users to provide feedback to improve the coach's accuracy and relevance. |
Additional fields
-
Additional Policy rules — Free-text field for any custom rules not covered by the toggles above (e.g. "If a client name is not found in the data, say 'I don't have data for [client name] — check the spelling or let me know their Fuse instance name.' Never attempt to guess or approximate.")
-
Additional Information — Contextual information the coach should know about its data sources, refresh dates, or limitations (e.g. "Data sources cover: customer health (Custify), feature page usage (last 180 days), open bugs (Jira/PD), and product feedback (Productboard, Userpilot, Pendo, Slack). Data was last refreshed March 2026.")
Tip: Start with Focus and Scope and Knowledge Limits enabled. These two settings alone will significantly improve coach quality. Add other policy controls as needed based on your use case.
Options
The Options sub-tab controls the chat experience — what users see alongside the coach's responses.
Prompt Library
When enabled, displays the prompt library for all connected contexts. This gives users pre-built prompts they can use to interact with the coach.
Source attribution and follow-up
These settings control how the coach surfaces its sources and encourages further interaction:
|
Setting |
What it does |
|---|---|
|
Show source content |
Displays the segments and files from the knowledge base that informed the coach's response. Choose between Compact list view (collapsed by default, users expand each type) or Card view (shown as a carousel below the response). |
|
Show speakers |
Shows speaker names from audio and video files used as sources. |
|
Show follow-up questions |
Displays AI-generated follow-up questions based on topics found in the source content. |
|
Show quick replies |
Displays clickable buttons when the coach suggests clear options or next steps. |
|
Show citations |
Shows inline citations within responses that link back to the source content in the knowledge base. |
Dynamic in-chat components
Enable or disable dynamic components that the coach can use to format its responses. Each component can be toggled on or off, and customised with specific guidance via the Edit Instructions button.
Available components:
|
Component |
Description |
|---|---|
|
Accordion |
Collapsible sections for organising content |
|
Bar Chart |
Compare values across categories |
|
Callout |
Highlight important information or tips |
|
Card |
Display content in a card layout with title and description |
|
Code Block |
Display formatted code snippets |
|
Feature Card |
Showcase features with icon, title, and description |
|
Line Chart |
Show trends over time |
|
Pie Chart |
Visualise proportions and distributions |
|
Steps |
Display sequential steps or instructions |
|
Tabs |
Organise content into tabbed sections |
Note: These components give the coach the ability to render rich, visual responses in the chat. Not all coaches will need all components — disable any that aren't relevant to your use case to keep responses focused.
LLM (model selection)
The LLM sub-tab lets you choose which AI model powers the coach. Different models have different strengths in terms of reasoning, speed, and response style.
Available models
|
Model |
Notes |
|---|---|
|
GPT 4.1 Recommended |
Great for everyday use. Handles most text tasks well — writing, answering questions, and coding. Good at remembering long conversations and following instructions carefully. |
|
GPT 4o |
Older generation LLM. Quick responses with good quality. Available until 31 May 2026 — we recommend switching to a newer alternative. |
|
GPT 5 |
Fast and capable for most tasks. Good value for solid performance without paying for the latest models. Available until 31 May 2026. |
|
GPT 5 Thinking |
Takes time on hard problems for thorough answers. Good value for careful reasoning. Available until 31 May 2026. |
|
GPT 5 Mini |
Smart but affordable. Good intelligence at lower cost. Faster than full GPT-5 while still being very capable. |
|
Claude Sonnet 4 |
Thoughtful and creative. Very good at writing and analysis. Known for careful, well-reasoned responses. Available until 31 May 2026. |
|
Gemini 2.5 Pro |
Top-tier reasoning and coding. Handles massive context (1M+ tokens) for big or complex projects. Available until 31 May 2026. |
|
GPT 5.1 |
Warm and conversational with a natural, human feel. Great for brainstorming, drafting, and tasks where tone and engagement matter. |
|
GPT 5.1 Thinking |
A more thorough version of GPT 5.1. Reasons more carefully behind the scenes, delivering more accurate answers on complex tasks. |
|
Claude Sonnet 4.5 |
Smart and reliable. A strong all-rounder for writing, analysis, and everyday professional tasks. |
|
GPT 5.2 |
Noticeably stronger at documents, long-context tasks, and multi-step projects. The reliable choice when quality and accuracy are essential. |
|
GPT 5.2 Thinking |
The most powerful option in the GPT 5.2 family. Applies deeper reasoning behind the scenes for more polished, accurate answers. |
|
Claude Sonnet 4.6 |
Thoughtful and creative. Very good at writing and analysis. Known for careful, well-reasoned responses and being the most creative. |
|
Claude Opus 4.6 |
Anthropic's most capable model. Exceptional at complex reasoning, multi-step analysis, and ambitious tasks. Best when quality matters more than speed. |
Things to keep in mind: Model updates happen periodically — when a model is updated, your coaches benefit from improvements automatically. Changing models doesn't change knowledge; it changes how the coach processes and communicates that knowledge. For most coaches, the recommended default (GPT 4.1) will work well.
Preview
The Preview sub-tab provides a live chat interface where you can test the coach before users see it. Use Preview to validate that the coach responds correctly, stays within scope, uses the right tone, and surfaces relevant knowledge base content.
Tip: Test with the kinds of questions your users will actually ask. Check whether responses are grounded in knowledge base content or falling back to generic AI responses. Iterate on the Purpose, Personality, and Policy settings until the coach behaves as expected.