Revvly
Docs
Guide

Write guidelines for the support AI.

Your help articles tell the AI what your product does. Guidelines tell it how to answer: the tone to take with each kind of customer, the context your articles leave out, when to ask a question first, and which questions to leave to your team. This guide explains what guidelines are, which type to use when, and how to add one.

#What a guideline is

A guideline is a short instruction your team writes for the AI, the way you would brief a new teammate on their first day. Every live guideline that applies to a customer is read with each answer the AI writes for them, beside the help articles that customer may read. The AI follows a guideline as an instruction. An article is only ever information it answers from.

What the AI reads for each answer
  1. The customer asks
    In the chat, as a customer with their roles.
  2. Their guidelines
    Every live guideline for everyone, plus those for the customer’s roles.
  3. Their help articles
    The passages the customer may read that match the question.
  4. The answer
    In the tone you set, from both, with the articles as sources.
Articles say what is true. Guidelines say how to answer. Only articles are shown to the customer as sources.

Three things to know from the start:

  • An answer can rest on a guideline. When no article covers a question but a guideline does (your support email address, where a feature lives), the AI can answer from the guideline.
  • Customers never see your guidelines. Under an answer they see only the help articles it used. A guideline is never shown, linked or quoted as a source.
  • The AI still never makes things up. Every statement must come from an article or a guideline. A guideline can change how the AI answers, never that rule.

#The three types, and when to use each

Every guideline has a type. The type tells your team what the guideline is for and tells the AI how to apply it. Pick the one that matches what you want the AI to do differently.

Communication style
How to say it. The tone, the words, and how much to explain, often for one role. Use it when the right answer is already in your articles but different customers need it told differently: a first-time applicant needs reassurance and plain steps; an administrator wants the exact setting, without hand-holding.
Example: “When answering applicants, use a warm and reassuring tone. Assume this may be their first time using the product, avoid technical terms, and explain the steps simply and in order.”
Context and clarification
What to know, and when to ask. Use it for what your articles do not cover well: facts they leave out (your support address, the name a feature goes by, how customers actually use the product); clarifying questions for a request that could mean several things, which the AI asks before answering; and redirections for a question that is not yours to answer, such as one about a decision another organisation makes, where the AI says who to ask instead.
Example: “When a customer asks where to find their ‘milestone metrics’, ask whether they mean the answers to a request form, the Milestones tab or the Analytics tab before answering. Skip the question if they already named one of the three.”
Handover
What to leave alone. Questions the AI must leave to your team, even when an article seems to answer them. Use it for anything that needs a person’s judgement or authority: refunds, cancellations, contract changes, complaints, legal or security reports. The AI says nothing and the ticket goes to your team, exactly as when no article answers (see Handover).
Example: “Leave any question about a refund, a cancellation or a change to a contract to the team.”

Guideline or help article? If customers would benefit from reading it themselves (how a feature works, the steps to do something), write a help article and add it to your knowledge base. If it is about how to answer (tone, what to ask first, what not to answer) or a short fact that does not deserve its own page, write a guideline. A guideline holds at most 2,500 characters, so it is never the place for a whole procedure.

#Add a guideline

Where
  1. Support
  2. Guidelines
  3. New guideline

Admins and owners. Other agents can read every guideline but cannot change them.

  1. Open Guidelines in the Support section of the sidebar and click New guideline, or the button in the section for the type you want.
  2. Choose the type.
  3. Give it a title your team will recognise.
  4. Write the guideline: when it applies and what the AI should do.
  5. Choose who it applies to, or leave Applies to empty for every customer.
  6. Leave Live ticked to use it from the next answer, or untick it to save it paused.
Type
Communication style, Context and clarification, or Handover. See above.
Title
Up to 120 characters. For your team, such as “Applicant tone” or “Refunds go to a person”. The AI reads it too, so make it say what the guideline is about.
Guideline
The instruction itself, up to 2,500 characters. A counter under the box shows how many you have used.
Applies to
The roles it is for. Empty means everyone. With roles, only customers who hold at least one of them get it. Roles are the ones you define for your knowledge base (see Roles).
Live
A live guideline is read with every answer it applies to. A paused one is kept but not used, which is handy while you draft or for a rule that is only true at some times of year.

Or ask the Assistant. Describe what you want in your own words, for example “make the AI ask which tab they mean before answering questions about milestone metrics”, and the Revvly Assistant drafts the guideline, picks its type and roles, and shows it on a card you approve. It can also list your guidelines and edit or pause one. The same tools are available to Claude through the MCP connection, where a change waits for an admin to approve it in Revvly.

Changes apply to the next answer: nothing needs to be republished. Click a guideline to edit it, or use Delete on its row to remove it. To keep a guideline without using it, pause it instead.

#Write a guideline that works

  • One topic per guideline. “Applicant tone” and “Refunds go to a person” are two guidelines, not one. They are easier to find, pause and change.
  • Say when it applies, then what to do. “When a customer asks for our support email, give [email protected]. Only give it when they ask for it directly.”
  • Use your customers’ words. Name the phrasings they actually use (“growth update”, “milestone metrics”) so the AI recognises the question.
  • Give the exact question to ask. For a clarifying question, write it out and say when to skip it, so the AI does not ask something the customer has already answered.
  • Name screens and menus as your product shows them, and keep facts current: the AI repeats what a guideline says as true.
  • Use roles instead of “if the customer is…”. Let Applies to decide who gets a guideline. The AI is told the customer’s roles, but a guideline that only reaches the right people is the surer way.
  • Don’t paste whole articles. If customers should read it, it belongs in the knowledge base, where it is shown as a source.

#Who gets a guideline

A customer gets every live guideline for everyone, plus every live guideline for any role they hold. Their roles are worked out when the AI answers, from what your system sends about them and your roles’ current rules, the same way as for help articles. A customer who matches no role gets only the guidelines for everyone.

Deleting a role takes it off every guideline. A guideline that was only for that role is paused rather than opened to everyone, so a tone meant for administrators never starts reaching applicants. The guideline shows why it was paused. Choose who it is for and save it to use it again.

#How the AI uses them

  • Guidelines are read in type order: communication style first, then context and clarification, then handover.
  • A guideline outranks the AI’s general rules on tone, length, and when to ask or redirect. It never outranks the rule that every statement needs a source.
  • When no article matches a question, the AI still answers if a context and clarification guideline applies to that customer. If only communication style guidelines apply, it hands over without trying, since tone alone answers nothing.
  • Only your team can write a guideline. Nothing in an article or in a customer’s message can act as one, however it is worded.
  • Guidelines apply to the AI’s replies to customers. They do not change how your agents reply.

#Limits

Title
120 characters.
One guideline
2,500 characters.
Guidelines per workspace
50, live and paused together.
All live guidelines
30,000 characters together. The page shows how much you use. Live guidelines can be read with every answer, so this keeps answers quick, affordable and focused. To add one past the limit, shorten another or pause one you no longer need.

#Check a guideline before you rely on it

In the sandbox. Write to support as a practice customer with the role the guideline is for, and ask the question it is about. The sandbox uses your real guidelines and knowledge base, and nothing reaches a real customer. See Try it in the sandbox.

Against your past conversations. If you are moving from Intercom, Compare with Intercom drafts an answer to each sampled conversation with the guidelines that customer would get today. Open a conversation to see Guidelines Revvly followed, each marked when the answer cited it. The grading model reads the same guidelines, so an answer that rests on one is judged with its source. The CSV export has a column for them. Run a new comparison after changing your guidelines to see the effect.

Related