You write the reply in the signal's detail pane, in the field pinned at the bottom. It is a single field that grows with the text.

If the responder agent has already prepared a text, the field is called Reply draft and holds a ready version. If there is no agent, the field is empty and called Write a reply manually.

Sending

The buttons below the field depend on the channel:

⚠️ Important: direct sending is not available on LinkedIn, Reddit, X, Bluesky and Hacker News yet - there we always prepare a draft that you copy and publish yourself.

If your channel account is not connected or the connection has expired, Telegrin shows a banner about it with a button that takes you to settings, and offers the text for copying.

Refining the text

There is a row of tools above the field. Four quick buttons apply in one click:

• Shorter
• Friendlier
• Add a question
• Drop the pitch

Next to them is the What should we change in the reply? field. Write your instruction in plain words ("mention that we work in Kyiv", "drop the mention of price") and press Enter - AI rewrites the text with your correction.

If the new version is worse, press restore the previous version - as many steps back as you like.

⚠️ Important: every refine costs 1 credit - the same as a full reply generation. That applies both to the quick buttons and to a correction sent with Enter. If you only need to change a couple of words, editing the text by hand right in the field is cheaper and faster.

What a reply costs

If the message did not go through

Telegrin shows a human-readable reason and returns the text to the field - nothing is lost. The most common reasons:

• the lead's privacy settings do not allow a first message;
• the lead has blocked your account;
• Telegram has temporarily limited the sending rate;
• the message is longer than 4096 characters - Telegram will not accept that much;
• the account connection dropped.

The card gets a Reply not sent marker: you can try again from the same pane.

If the agent sent the reply itself

When automatic sending is on for the responder, there is no reply box - instead you see a line saying the agent completed the step automatically. The text that was sent stays visible in the conversation.

Continuing the conversation after the first reply

After the reply, the conversation continues in the same pane: write in the field and send it like a normal message. The signal's verdict does not change, and the exchange is stored on the lead's card.

How to make replies better

If the texts are systematically off, the problem is almost never the individual reply:

1. The company brain. The most common cause of generic replies is empty or overly short Product and Advantages sections. AI does not invent facts about your business: what is not in the brain will not be in the reply.
2. The responder prompt. A prompt is an instruction, not a template: the agent writes its own short text every time (usually 2-4 sentences with one call to action). So it paraphrases and compresses, and can drop a specific wording.
3. To make sure a piece of text ends up in the reply, write it in the prompt as an imperative: "always add the link example.com at the end, word for word" works, "we have a website example.com" does not. Descriptive sentences are treated by AI as background it may ignore.
4. Give an example of the reply you want right in the prompt, in quotation marks - one or two samples work better than an abstract description of tone.

ℹ️ Note: the language of the reply is not set by the prompt. The agent replies in the language of the post or of the lead's message - that rule outranks both the prompt and the brain. All you can change is the fallback for cases where the language could not be determined.