Skip to content

Write agent instructions and choose its actions

An agent does what its Instructions say, and nothing past replying unless one of its actions is switched on. The agent editor, on the AI Agents tab of Inbox, keeps its settings in sections down the left: Overview, Messages it sends, How it talks, What it knows, What it can do, Hand-overs, Hours and limits, Lessons and Advanced. On a phone they’re a row of tabs. Type into Find a setting above them to jump straight to one, like reminders or quiet hours.

To brief the agent:

  1. Open Inbox, then AI Agents, and press an agent’s Edit.
  2. Open How it talks and type into the Instructions box, in short headed sections.
  3. Use Add prompt templates to drop a ready-written section in.
  4. Press Optimize if you’d like it rewritten more clearly.

Under the box, Learn how to write this opens How to write instructions, which suggests five parts in order: who the agent is, what its first reply does, the steps of a good conversation, what it must never do, and then a test.

Add prompt templates offers Greeting, Qualifying questions, Booking a call, Follow-up on silence, Escalation and Boundaries. Whichever you pick lands at the end, after what you’ve already written.

Optimize keeps every fact, rule, name and number and rewrites the wording step by step. It refuses a rewrite that drops the customer’s name or comes back much shorter than yours, and leaves your text as it was. The message after a rewrite offers Undo.

The @ button under the box types the name of the advisor or the business this agent speaks for. The {} button inserts Customer's first name, which goes in as $contact.firstname.

Before each reply, the agent reads it as the first word of the customer’s name. A conversation with no real name, such as one labelled with a phone number, gets “there” instead, so the greeting reads “Hi there!”.

What it can do lists what the agent may do to a chat or a card, in 2 groups. Each group heading says how many of its actions are on. Open the info mark beside an action’s name for a one-line description.

Under In the chat:

  • Close conversations closes the thread with a short summary and whether it ended won, lost or neither. The summary goes under the thread’s Details, in On this chat.
  • Add comments leaves notes for you under the thread’s Details, which the customer never sees.
  • Trigger workflows starts one of your published workflows for this contact.

Under On their card:

  • Update lifecycle stages moves the lead between new, replied, booked, won and lost.
  • Update contact fields fills an empty field on the contact with a detail the customer wrote, such as a name or an email. Anything else waits under AI agent on the contact, for you to Use this or Dismiss. See Review what the AI agent heard.
  • Update tags adds or removes tags on the contact.
  • Take referrals puts a customer someone passes to you on your board, and sends them the first message.

Hand-overs holds Assign to you or your team, which stops the agent and hands the conversation to a person, with a reason. Change where handovers go, under Who gets told on its card, sends that to your Lark group, a WhatsApp, or both.

Messages it sends holds the actions that write to the lead on their own:

  • Book appointments offers your free times and books one on your Pick a Time page.
  • Remind about appointments messages the lead before a booked appointment, on a cadence you set. Each reminder has an Exact words box. What you type there goes out as written, with [Name], [Date], [Time] and [Link] filled in for each booking.
  • Ask for a Google review asks happy customers for a review once their job is done, and hands unhappy ones to you.

At the top of Messages it sends, In the order a lead gets them lists every message it sends without being asked, with when it goes. Each line says Exact words, Agent writes or Off, and pressing a line opens its setting.

Switch an action on and a box opens under it headed When and how should this action be performed?. Write when the agent should use it. Each box has its own Optimize, and most also offer Add prompt templates written for that action. Under Ask for a Google review, the box is headed Anything to say or avoid when asking? instead.

Follow-ups sit in Messages it sends. A new agent starts with them off.

Follow up a chat that goes quiet writes to a lead whose reply to our last message never came, in a chat the agent is handling. Its slider has 3 stops: Off, Once and Twice. Each follow-up goes a day after our last message, and then the agent waits for their reply. Note that every unanswered message counts against the WhatsApp number, which is why the card says This can get the number blocked once the slider is past Off.

On a financial adviser’s agent, follow-ups also reach people who never replied to your first message, and the card says so. A business’s agent never follows up somebody who hasn’t replied. When they go out sets the hours a follow-up may go out, Singapore time, and starts at 09:00 to 21:00. Only when our last message asked them something is on by default, and Only chats active in the last limits it to chats from the last 7 days.

When they say they’ll get back to you, the agent waits for them instead. When they say they will get back to us, check in after sets how many days, and starts at 3. A day they name themselves, like after Chinese New Year, wins. That check-in still goes with the slider at Off.

Check in on quiet cards, lower in the same section, checks in with leads whose card has gone quiet in a pipeline column.

Overview opens with How it behaves, one line for each rule below, and pressing a line takes you to its setting. Wait time, Quiet hours and who it answers sit in Hours and limits. The reply limit and what it does when it’s unsure sit in Hand-overs.

Wait time is how long it waits after their last message before replying, so one reply covers everything they sent. It’s 3 seconds by default, and it takes 0 to 60. It waits a little longer while they keep typing, but never more than a minute.

Quiet hours takes a From and an Until, in Singapore time. A new agent starts at 23:00 to 09:00. A message that lands between them is answered when they end. Switch on Still answer someone chatting now and it replies at once to anyone who writes back within an hour of our last message.

Stop after this many replies hands the chat to you. It’s 50 by default, and it takes 1 to 100. The count starts again when you press Back to the agent, and the line over the chat shows it, such as 3 of 50 replies used.

Answer people who message first decides who the agent answers. Off, it answers only people you reached out to, like leads from your ads, and anyone else who messages the number goes to you. It’s off by default for a financial adviser’s agent and on for a business’s.

When it is not sure is on by default. The agent then says nothing and hands the chat to you rather than guess. Languages, in How it talks, names the languages it handles, and it’s told to answer in the one the customer writes in.

Advanced holds Room for one reply, and a district admin also sees Model, Effort and House rules there.

Every agent is also given rules after yours, and your instructions can add to them but not remove them. It’s told never to give financial, investment, insurance, tax or legal advice, never to quote prices, premiums or returns, and never to recommend a specific product, fund or provider. An agent an admin sets to Business: quotes its own prices may quote prices written in its instructions or knowledge instead.

It writes as you, in the first person, and never says it’s an AI. If somebody asks outright whether they’re talking to a person, it hands the chat to you. It’s also told never to ask for an NRIC, bank details, card numbers, passwords or documents. A complaint, a mention of a regulator or PDPA, or a question its knowledge can’t answer should end with the conversation handed to you.

Is there a limit to how long the instructions can be?

Yes. Instructions stop at 40,000 characters, and the editor won’t save them until you shorten them. Each action’s guideline keeps its first 4,000. The count under the box shows how many you’ve used.

What happens to a detail my book has no column for?

It waits on the contact, under AI agent and Not recorded. Use this makes the column and fills it in.

Will the agent change a value I typed?

No. It fills empty fields, or corrects a value it wrote itself, and only with the customer’s own words. A different value waits under AI agent on the contact. Type the same value in yourself and the suggestion clears.

Do changes go live as I type?

No. Nothing is saved until you press Save draft or Publish. On a published agent the button reads Publish changes, and that save is live straight away.

Does the template already switch actions on?

Yes. Receptionist starts with assigning and closing, with follow-ups set to Off. Sales agent has those plus stages, contact fields and comments, while Support agent has them plus comments.