AI for comment keywords and DMs: 5 prompts to copy

One good DM is easy. A keyword, a DM, five public replies and a follow-up for every post, every week, is where most people quietly stop. That repetitive writing is exactly what AI assistants like Claude and ChatGPT do well.

What they cannot do is know your prices, your opening hours or how you talk to customers. This guide shows how to give them that context, five prompts you can paste today, and where a person has to stay in the loop.

In short:

• AI is a fast first-draft writer for keywords, DMs and comment replies.
• The quality depends on the brief you give it, so write that once and reuse it.
• It will invent prices, dates and links if you let it. Forbid that in the brief.
• A person reads and approves every message before a customer sees it.

Start with a brief, not a prompt

Every new chat should begin with a short description of your business. Write it once, save it in a note, paste it each time:

I run [business] in [city, or online]. I sell [product or service] for [price]. My customers are [who]. My tone is [friendly / professional / playful]. I use comment-to-DM on Instagram: people comment a keyword under a post and get one automatic DM. Rules for every DM: two short lines at most, include the link placeholder [link], end with one easy question, no more than one emoji. Never make up prices, dates, stock or links. If you do not know something, write a placeholder in square brackets.

The last two sentences do the most work. Without them, the model fills gaps with confident guesses.

One detail worth adding if your audience is on Instagram: the DM can greet people with a name variable, but Instagram does not share a commenter's name, so the greeting reads "Hey there". Tell the AI to write openers that work either way.

The 5 prompts

1. Keyword ideas

Suggest 10 comment keywords for a post about [offer]. Each must be a single short English word that tells people what they will receive. Avoid vague words like INFO or YES, avoid brand names and numbers, and avoid words people use in normal comments. For each keyword, write the opening line of the caption that asks for it.

Then choose. Prefer the word that names the reward. Caption CTA formulas helps you judge the opening lines it gives you.

2. The first DM

Write 3 versions of the DM a person receives after commenting [KEYWORD]. First line: the answer to [their question] plus [link]. Second line: one easy question that moves them to the next step. Under 40 words each.

This is the two-part structure from sales DM templates: give the thing, then ask one question.

3. Public replies

Write 6 short public replies to a comment, telling the person I've sent them a DM. Each under 8 words, all different, warm, no hashtags, one emoji at most.

4. A critique of your current DM

Here is my DM: "[paste]". Pretend you just commented [KEYWORD]. What is unclear? What would stop you from answering? Rewrite it as answer plus one question, and explain each change in one line.

This one is most useful after a week of real replies, when you know what people actually asked.

5. A follow-up you will send yourself

A customer asked about [product] yesterday, got the details, and did not buy. Write 3 short follow-ups I can send personally today. No pressure, no discount, one question each.

The AI writes this one, you send it. An automation cannot chase someone who stayed silent.

What AI does, and what stays with you

Copy-paste setup: what an edited result looks like

Say a small tea shop runs prompts 1 to 3. The AI suggests TEA, BLEND and AUTUMN. The owner picks BLEND, because it says what you receive and nobody types it by accident. The AI also wrote "steeps perfectly in 3 minutes", which nobody has tested, so that line goes.

• Caption, line one: "Three new autumn blends are in. Comment BLEND and I'll DM you the list."
• Trigger: comment keyword BLEND, on this post only.
• First DM: "Hey {{first_name}}, the new blends are smoked pear, chai rooibos and ginger mint. All three are here: https://yourshop.com/autumn. Loose leaf or tea bags?"
• Public reply variants: "Sent, check your DMs", "The list is in your inbox {{username}}", "Just messaged you".
• Tag: blend-autumn.

The question at the end works as a sorter. If you want answers to trigger something, ask for a specific word ("Reply LOOSE or BAGS") and give each word its own DM keyword flow.

Can AI work inside your tools, not just a chat window?

Yes, if the tool has an MCP server. Chat33 has one, and when you connect Claude or ChatGPT to it, the assistant can:

• schedule posts to your connected channels and read post analytics,
• list your DM conversations, read them and send replies,
• read comments on your Instagram and Facebook posts and reply publicly or privately,
• create new DM automation flows and switch them on or off.

The details, and setup guides for each assistant, are on the MCP page and the agents page.

Two facts matter for how you use this. A flow the assistant creates starts as a draft and does nothing until it is switched to active, so you have a natural place to review it. A DM reply sent through the assistant, on the other hand, goes out right away, and only inside the 24-hour messaging window. So my rule is simple: AI drafts, a person approves, then it goes out. Ask the assistant to show you the text first and only then tell it to send.

I have built software for about 25 years, and the tools that last make a person faster instead of removing them. AI fits that pattern well when you keep the approval step.

A one-week plan

1. Write your business brief and save it.
2. Pick one post that gets the same question every day.
3. Run prompts 1, 2 and 3, choose, edit, check every fact.
4. Build the flow, then test it from a second account.
5. After a week, run prompt 4 on the DM using the real replies.

Before you launch, compare your draft with the most common DM automation mistakes. AI produces several of them on its own, especially the polite "thanks for your comment" opener.

Instagram comment-to-DM in Chat33

Keyword triggers, rotating public replies, tags and tracked links, plus an MCP server for Claude and ChatGPT. Also works on Facebook Pages.

FAQ

Is it OK to send DMs written by AI?

Yes, as long as a person reviews them. The risk is not the wording, it is invented facts or promises. Check prices, dates and links before anything goes live.

Should I paste customer data into ChatGPT or Claude to write templates?

No. You do not need real names, phone numbers or order details to write a template. Use placeholders and keep personal data out of the chat.

Can Claude or ChatGPT reply to my Instagram DMs automatically?

Through Chat33's MCP server, an assistant can read conversations and send replies, but only when you ask it to, and only within the 24-hour window after the person's last message. Keyword flows handle the instant replies. The assistant is best used for drafting and reviewing with you.

Which AI is better for writing DMs, Claude or ChatGPT?

Both write good drafts when given a clear brief. The brief and your editing matter more than the model. Pick the one you already use and keep the same brief in both.