Fin Guidance allows you to fine-tune how Fin responds to customers, ensuring its answers are accurate, on-brand, and aligned with your support policies. To get the best results, follow these best practices when writing prompts.
Fin Guidance best practices
Start with the outcome in mind
Before writing guidance, consider the specific result you want to achieve. Work backward from this goal to create clear, actionable instructions.
❌ Bad example (vague and ineffective):
“Make sure Fin understands our different product types before answering.”
✅ Good example (clear and structured):
“If a customer asks about the ‘search’ feature, first ask which product they are using before responding. Then, provide product-specific instructions based on their answer.”
By structuring guidance with clear steps and conditions, Fin can apply it consistently and correctly.
Speak directly to Fin in your guidance
When writing guidance, avoid referring to Fin in the third person or commenting on
how it should modify its responses. Instead, write as if you are speaking directly to
Fin, telling it exactly what to do.
❌ Bad example (third-person and indirect):
"If the Al answer tells the customer to uninstall and reinstall the app, then rewrite the
answer to remove that information, as reinstalling is not a valid troubleshooting
step."
✅ Good example (direct and actionable):
"Never tell the customer to uninstall and reinstall the app. Reinstalling is never a valid
troubleshooting step, and Fin should never communicate this to the customer."
Use simple and precise language
Ambiguous or overly complex guidance can lead to inconsistent responses. Write as if you’re training a new support agent—be direct, specific, and easy to understand.
❌ Bad example (unclear and open to interpretation):
“Fin should be professional but also friendly, keeping responses engaging.”
✅ Good example (defined and actionable):
“Fin should use a professional yet approachable tone. Keep responses concise, avoid jargon, and use reassuring language when addressing customer frustrations (e.g., ‘I understand how that can be frustrating. Here’s how we can resolve it’).”
Provide context and concrete examples
Fin performs best when it understands when and how to apply guidance. Use words like “if,” “when,” and “then” to define conditions, and include clear examples.
❌ Bad example (lacks context):
“When customers ask about pricing, make sure Fin answers correctly.”
✅ Good example (context-driven and detailed):
“If a customer asks about pricing, first check if they mention a specific plan. If they don’t, ask which plan they’re interested in before providing details. Always refer to prices as ‘starting at [lowest tier price]’ unless the customer specifies a plan.”
Create separate, focused guidance
Each piece of guidance should address a single objective. Avoid mixing multiple instructions, as this can make it harder for Fin to apply them correctly.
❌ Bad example (too broad and unfocused):
“Fin should use a friendly tone, clarify questions before answering, and escalate billing issues to an agent.”
✅ Good example (one clear purpose per entry):
Communication style guidance: “Fin should maintain a warm, friendly tone by using positive language and avoiding overly formal phrases.”
Context and clarification guidance: “If a customer asks about refunds but does not specify a purchase date, ask for the date before proceeding.”
Handover and escalation guidance: “If a customer mentions ‘billing error’ or ‘overcharge,’ immediately escalate the conversation to a human agent.”
Use article titles to guide responses
If you want Fin to reference specific source content in its response, you can mention the title of the source in your Guidance.
Important caveats:
Guidance does not influence which articles are retrieved by Fin.
If an article wasn’t retrieved based on the customer’s original query, Guidance can’t make Fin use it.
Guidance can’t add extra context to the customer’s query in order to retrieve different or additional content.
Guidance only helps Fin generate a better response using the content it's already retrieved.
So, your Guidance should be based on the customer’s original query and the content that was likely retrieved as a result.
❌ Bad example (tells Fin which content to use):
"If the customer says 'upgrade' use information from the source titled 'Subscription Management' to form your response."
✅ Good example (helps Fin respond if the relevant content is retrieved):
“If a customer asks about upgrading their plan, which can be found in the source titled ‘Subscription Management,’ then make sure you highlight where they can do this and include all the steps to upgrade a plan in your response.”
Determine when to use audience targeting vs. user attributes
If you’re deciding whether or not your Fin Guidance should apply based on a user’s data, it’s usually best to use the Audience targeting feature. Targeting an audience completely hides any non-matching guidance from Fin, which means there’s no risk of confusion or duplication.
For example:
If you only want the guidance to apply when a user is on the “Pro” plan, create an audience for "Plan is Pro", then select this audience when creating your guidance. Fin will only see that guidance when it’s relevant.
On the other hand, if you want Fin to reference the actual value of a user attribute in its response, you should include that attribute directly in your guidance.
For example:
“Address the user by name:
{user_name}.”“Use formal pronouns if
{user_pronoun_formality}is ‘formal’.”“Mention features available in
{user_plan}when describing functionality.”
Quick rule of thumb:
Use audience rules if you’re filtering when the guidance should apply.
Use attributes in the guidance itself if you’re customizing the response content based on user data.
Continuously refine your guidance
Think of Fin Guidance as an ongoing process. Start with essential instructions and improve them over time based on real interactions and performance metrics.
Test Fin’s responses in Preview Mode to ensure the guidance works as expected.
Regularly review and refine guidance based on metrics and customer feedback.
Remove or update conflicting or redundant guidance to keep Fin’s behavior consistent.
Tip: If you’re unsure how to write a prompt for a specific use case, try using a writing tool like Claude AI or ChatGPT. Describe the scenario and ask it to generate a clear, AI-friendly prompt that’s easy to interpret.
Fin Guidance examples
Below you will find some examples of guidance. You can use these as inspiration or adapt these for your specific needs.
Communication style
Specific vocabulary and terminology Fin should use.
We offer Free, Pro, and Enterprise plans. They should be capitalized and referred to as “plans” rather than “subscriptions”.
Always answer in British English.
Use plain language, avoiding jargon or buzzwords. For example, use "easy" instead of "frictionless".
If the user is frustrated, show empathy and use calming language to let them know that you care about helping them out—for example, “I understand this is frustrating, and I apologize. I am going to work through this with you to find a suitable solution”. Feel free to use different variations of this example.
Answers should be always readable and concise: sentences should be short, there should be at most 1-2 sentences per paragraph and no more than 100 words per answer (unless absolutely necessary). Simply break paragraphs into sentences with a newline without other formatting or content changes. Note that this guideline only applies to regular textual paragraphs, it does NOT apply to code, bullet points, list, and other structured markdown.
Never ask the user to contact <company name> through email as they're already contacting <company name>'s customer support through the chat messenger. For example, remove passages such as "please email us at <company email>" from any AI message. In those cases, tell the user instead that if they need further help, they should request human assistance by clicking on "Talk to a person 👤".
We are currently in the holiday/new year period. End interactions with a brief, inclusive seasonal greeting that fits the customer's language and region when clear from the conversation. Keep greetings short and at the end of responses. Avoid holiday wishes when handling urgent issues or customer complaints.
Context and clarification
Follow-up questions Fin should ask, to ensure accurate answers.
If a user is having an issue completing a purchase, ask if they are trying to use bank transfer or a credit card.
If a user asks what features they have access to, first clarify which plan they are on.
If a user reports that they are seeing a 502 error, let them know that we are aware of the problem and are currently investigating as a matter of priority. They can follow for updates on: https://www.intercomstatus.com/
Do not assume what mobile app the user has. Always ask if they are an iOS or Android user when they report an error. Each app has a different subset of features, so it’s important to understand this first.
When a user reports a bug, first ask for the following information before continuing troubleshooting: 1) what device are you using? 2) URL where the bug is occurring; 3) when was the bug first noticed.
If the customer inquires about delivery options, check their {{country}} and provide a response that aligns with the delivery options available for that specific country.
Handover and escalation
Moments when Fin should route customers straight to your team.
If a user asks about cancelling their plan, always hand the conversation over to a human agent.
If the user mentions any of the following tax-related terms, always hand the conversation over to a human agent: tax planning, tax consulting, tax advice.
If a user requests a task that requires agent action, always hand the conversation over to a human agent immediately. These tasks include: Exporting the analytics dashboard as a CSV, Generating a code snippet for camera installation, Updating a compliance document
When a patient requests to book an appointment with Dr. Mills, always hand the conversation over to a human agent.
If the customer’s {Pricing Plan Name} is "Pro", always hand the conversation over to a human agent.