Prompting Your AI Colleague

When you hire a colleague from the catalog, the wording, escalation rules, tone, and edge-case handling are already written by us and maintained over time. You only fill in the Your details form on the Instructions tab — your company name, check-in / check-out times, access instructions, and similar facts.

Behind the scenes the colleague combines three things on every conversation:

1. Your details — the structured answers you filled in.
2. Live reservation and listing data — fetched fresh each conversation and matched to the guest by phone number or message thread. You don't paste this in; the colleague reads it automatically.
3. Your knowledge base — the articles you've given it access to.

A fully custom, hand-written prompt is an Enterprise-only override. On those workspaces an "override with a custom prompt" option appears at the bottom of the Instructions tab; it swaps the managed playbook for a raw prompt you write and publish yourself.

The sections below are the guidance that applies if and when you write a custom prompt.

The single biggest difference between a usable custom prompt and a great one is the number of concrete situations it anticipates. Don't stop at "answer guest questions" — list the specific scenarios that come up in your business.

For a guest-facing colleague that might include:

• Late-night arrival on the arrival date
• Early check-in request before 3 PM
• Guest asking for the WiFi password
• Guest unsure how to find the building entrance
• Guest asking about parking when the listing has none
• Refusal to fill the check-in form / share ID
• Mechanical lockout (broken key, code not working)
• Question about cancellation policy without cancelling
• Question about pet policy when no pet info is in the listing
• Guest asking to bring extra people beyond the booked headcount

For each scenario, write what the colleague should say and what it should not say. Include the wording you would use yourself if you picked up the phone — the colleague picks up tone from your examples.

Build a custom prompt in layers, from static to dynamic:

1. Static knowledge. Things that don't change per conversation — your check-in hours, parking rules, cancellation policy, brand voice, what you do and don't handle. This is the bulk of the prompt at the start.
2. Per-reservation context. Reservation-specific variables: who the guest is, their arrival date, what they booked, which channel. These are filled in automatically at message time using template variables like {{reservation.guest_first_name}} and {{reservation.arrival_date}}. The custom-prompt editor lists the available tokens for you.
3. Per-caller context (CRM, other systems). Eventually you can pipe in information about the specific person — their record, past tickets, lifetime value — to make answers situational rather than generic. Add these after the first two layers are solid.

Counter-intuitively, longer custom prompts often produce better answers — as long as the extra content is relevant. The model uses your wording, your examples, and your edge cases as anchors. Vague prompts force it back on generic language.

What "relevant" looks like:

• Specific scenarios from your day-to-day operations
• Example phrasings you actually use with customers
• Edge cases and what to do about them
• What NOT to say (e.g. never promise refunds, never share off-platform contact)
• Brand voice cues (formal vs casual, language preference)

What padding looks like (avoid):

• Long marketing copy that doesn't change behaviour
• Generic AI-assistant boilerplate ("be helpful and concise" repeated five times)
• Listing every possible variable when only a few apply
• Repeating the same rule in three different wordings

The reliable way to make a custom prompt better over time is the Tests tab — a regression suite that replays real conversations against your current prompt. When the colleague gets something wrong, you capture that exact moment as a test, use Fix with AI to propose a prompt change that makes it pass without breaking the others, review it, and publish.

• Approve the proposed changes that fix real issues
• Reject the ones that don't fit your voice or business rules
• Proposed changes are never applied to the live prompt behind your back — you review and publish

See Testing Your AI Colleague for the full loop. Over a month or two, this makes the prompt substantially richer without you having to predict everything up front.

• One-sentence prompts. "Answer guest questions" produces a generic colleague that fails on every nuance. Aim for at least one section per major scenario.
• Reaching for a custom prompt too soon. If the managed role nearly fits, filling in more detail and knowledge usually beats hand-writing a prompt from scratch.
• Setting it and forgetting it. A prompt that hasn't been touched in three months is missing three months of edge cases.
• Hardcoding values that should be variables. If you write "the apartment is on the 3rd floor" in the prompt, the colleague says that for every listing. Use {{listing.floor}} instead.
• No escalation rules. Always include explicit instructions for when to hand off to a human (complaint, emergency, request to speak to a manager).
• Inventing what you don't know. Tell the colleague to say "I'm not sure, our team will check and get back to you" rather than guessing. A colleague that guesses politely is worse than one that admits uncertainty.

• Hiring an AI Colleague — hire your first colleague.
• The AI Colleague Catalog — the ready-made roles you hire from.
• How Your Colleague Gets Its Knowledge — knowledge base access and live data.
• Colleague Triggers & Routing — how messages and calls reach the right colleague.