Drafts and templates

This page is about two ways to answer messages with help: a draft your agent prepares for you, and a template that answers by itself. Use it when you want help with replies and need to decide how much to hand over. By the end, you will be able to ask for drafts you approve one by one, and write an auto-reply template with your own text and, if you want, a part written by a model.

Two terms matter here:

• A draft is a reply your agent writes and shows you. Nothing is sent until you approve the exact text and the chat it goes to.
• A template is the text of an auto-reply rule. The background server (tg serve or max serve) fills it in and sends it on its own, without asking you, but only to people you allow and only after you turn sending on. Templates use Liquid, a small language for text with placeholders such as a sender's name.

Drafts: the agent writes, you approveCopy Anchor Link

You need a connected account and an agent. The agent reads the conversation, writes a reply and shows it to you. An agent connected through MCP is told that a draft, or "we should reply", is not a request to send: it may send only the exact text you asked for, in the exact chat. For any agent, what actually stops a send is the profile's permissions: a refusal, such as a read-only profile or a recipient that is not allowed, is final.

The examples use Telegram. For MAX, write "Use max CLI" instead of "Use tg CLI".

Reply in one chat:

Use tg CLI. Read the recent messages in my chat with Anna and draft a reply to her last question.
Keep it short and use only facts from our conversation. Show me the draft; don't send it.

Answer questions that are still waiting for you:

Use tg CLI. List the questions people asked me this week that I have not answered, oldest first.
For each one, show the chat and the message, and draft a short answer. Send nothing and mark
nothing as read.

Remind people of what they promised:

Use tg CLI. Review my chats since Monday: what I owe, what others owe me and what needs clarifying.
Draft at most five reminders, each with its chat and text. Show them to me; send none.

When a draft is right, approve it with the exact text and recipient, for example "Send exactly that draft to Anna's direct chat." If you change the wording, give the full final text. Your permissions still apply: a profile can require confirmation or forbid sending, and a request does not bypass that. More examples are in how to phrase requests.

If your AI app is connected through MCP, the tool also offers ready-made prompts for replying in a chat and for reviewing commitments. They follow the same rule: drafts first, a send only after you approve.

Templates: rules that answer by themselvesCopy Anchor Link

An auto-reply rule says when to answer, for example outside working hours, and its template says what to send. Before any reply goes out:

• the sender must be among the people you allow to receive auto-replies;
• sending must be allowed in the profile's permissions with permissions.replies.send allow; ask counts as no, because the background server has nobody to ask;
• every rule has a limit per chat and per person, such as one reply per chat in 12 hours.

replies test shows what the rules would have answered in your saved messages. It sends nothing.

What a template can useCopy Anchor Link

Standard Liquid filters work, such as default for an empty value and date for a time: {{ now | date: "%H:%M" }}. Conditions such as {% if chat.kind == "dialog" %} work too.

The incoming message's text is never a template variable. An unknown variable or filter makes the template invalid, and then no reply is sent. A template cannot read files, and its time, memory and length are limited.

ExamplesCopy Anchor Link

An away message:

Thanks, {{ sender.firstName | default: "there" }}! I'm away until Monday and will reply then.

A reply outside working hours. The hours belong to the rule, not to the template: set them with replies edit and the options --outside, --days and --timezone. Then now is in that time zone:

Hi {{ sender.firstName | default: "there" }}, it's {{ now | date: "%H:%M" }} here. I answer on weekdays from 9:00 to 19:00.

A part written by a model, with your own text as a fallback:

Thanks, {{ sender.firstName | default: "there" }}!
{% ai %}Briefly confirm you got the message; I will answer tomorrow.{% else %}I will answer tomorrow.{% endai %}

A part written by a modelCopy Anchor Link

A model may write only what is inside {% ai %}…{% endai %}; everything outside it is your text. The block's body is an instruction to the model. The incoming message goes to the model separately, as data, never as part of the instruction. The model's answer is used as plain text.

The text after {% else %} is the fallback. It is sent when no model is set up, when there is no consent, when the call fails, or when the model's answer is refused because it is empty, too long or repeats the incoming message. Without an else, that message gets no reply and the reason is recorded. Blocks cannot be nested.

Using a model needs its own consent, separate from the right to send. Consent lets incoming message data go to the model provider you chose, for the whole profile:

tg replies consents show
tg replies consents grant
tg replies consents revoke
max replies consents show
max replies consents grant
max replies consents revoke

replies consents deny with a chat ID keeps that chat's messages away from the model. A new provider or endpoint needs consent again. replies test --ai sends saved messages to the model so you can see its text; it still sends nothing to the messenger.

Set it upCopy Anchor Link

Start with drafts: they need nothing beyond a connected account. To write rules, follow Telegram auto-replies or MAX auto-replies, which cover creating a rule, choosing who may receive replies, the model settings and how to pause everything at once.