Posts, DMs and a content plan without an SMM manager: what do you set up so it all runs itself?

Of everything an SMM manager does, AI handles four routine things well, and none of them is strategy. Topics for the day, copy for each platform, scheduled publishing, replies in DMs. That is where most of the week goes, and that is exactly what you can set up once so it runs on its own, with no days off and no salary.

We built this kind of setup for our own channel. Configuring it took about six hours, and it covers what used to be a week of manual work. Below, step by step: which service sits in each link, what to configure in it, and where the free tier ends.

Which parts of an SMM manager's job AI takes, and which it doesn't

Let's start with money, because every conversation about replacing someone starts there.

The list of duties in the same MGUTU review mixes two different jobs into one role. The first is a production line: making content to plan, formatting posts, replying to comments. The second is decisions: strategy, the content plan as an idea, analytics, budget. The production line repeats every day and is easy to describe with rules, so that is what a chain of AI services takes over. Decisions stay with a person, and there is a separate section on that at the end.

If you want the whole picture first, not just SMM, we have an overview of eight digital agents that replace entire departments. Content there is handled by the ContentFlow agent, and this article is essentially its detailed build.

1. 1

   Topics and content plan

   A Claude agent runs on a schedule, finds topics, scores them and sends the list to Telegram. A person picks one.

2. 2

   Copy

   For the chosen topic, AI writes a post for each platform. A person edits and approves it; nothing goes out without approval.

3. 3

   Schedule

   Upload-Post spreads finished videos and posts across platforms into queue slots.

4. 4

   DMs

   ChatPlace replies on Instagram to anyone who comments a keyword under a reel: a greeting, a follow check, a link to the material.

Step 1. Content plan: a scheduled agent sends topics to Telegram

The morning topic list comes from a Claude agent that starts by itself. Claude Code has routines for this: a saved prompt, a repository and connected services that run on a schedule in Anthropic's cloud. Your laptop can be closed. You create them at claude.ai/code/routines or with the /schedule command in Claude Code.

What to configure:

1. A Telegram bot. You create it through @BotFather, which gives you a token. The bot only needs to send you the list and receive your choice.
2. The prompt. It has to be self-contained: the agent works without you and won't ask follow-up questions. Describe where to look for topics, how to score them and in what form to send them.
3. The schedule. The "daily" or "weekdays" preset. The docs advise against the very top of the hour: a run at 9:00 can be a few minutes late, while one at 9:07 arrives on time.
4. Network. By default the cloud environment only allows an allowlist of domains, and Telegram isn't on it. In the environment settings, switch access to Custom and add api.telegram.org. Put the bot token in the environment's variables. Anyone who uses that environment can read them, so keep a separate environment for the bot and don't share it.

Every morning, find 5 topics for posts in our channel about AI for business.
Sources: vendor blogs, changelogs, news from the last 24 hours.
Drop anything we already covered in the last 30 days.
For each topic: a headline, one line on why it matters to our
readers, and a link to the primary source.
Send the list to the Telegram bot with a pick button under each topic.
Don't publish anything yourself.

The ban on publishing sits in the last line for a reason. A scheduled agent runs unattended, and a rule like that is more reliable written out explicitly than left to the model's common sense.

Someone has to receive the button press. A routine starts on a schedule and won't hear the tap by itself, so the bot needs a handler: a workflow in n8n or Make, or a small cloud function. On a tap, it calls the routine's API trigger with the chosen topic. Keep in mind that by default a routine treats text arriving through the trigger as untrusted: say explicitly in its prompt that the topic should be taken from that text.

Step 2. Post copy for each platform

You tap the button under a topic, and AI writes the post. Not one post for every platform but a separate one for each: a reel caption, a Threads post, a shorter text for X, a YouTube description with keywords.

Copy quality here depends on what you give the model. Three things worth writing down once and attaching to every request:

• voice: five to ten of your best posts as samples, plus a list of words you never use;
• platform rules: length, hashtags, where a link works and where it doesn't;
• don'ts: what you never promise and what you don't write about.

These rules are easy to keep as a skill that kicks in on its own whenever the task is about posts. We showed how that works in our guide on Claude skills that work for you.

Then comes a person: you read the copy, edit it, approve it. At our end a human always approves publishing, and we consider that a required part of the setup, not a temporary measure.

What to configure:

1. Where generation runs. The simplest option is the same routine: the button handler calls its API trigger with the topic, and it writes the copy. If the post prompt differs a lot from the morning one, give the copy its own routine with its own trigger.
2. Where the draft arrives. The same Telegram bot: copy for each platform with an "Approve" button underneath.
3. What "Approve" does. The handler catches the tap and starts the routine again with the approved copy. The routine has Upload-Post's MCP server connected and uses it to put the post into the queue. More on that in step 3.

Step 3. Scheduled autoposting to Instagram, YouTube, X and Threads

The approved post and the finished video go to Upload-Post. It's a service with one API for about twenty social networks: it adjusts format, size and caption length for each platform itself.

What to configure:

1. Profiles. Connect your platform accounts. The free plan allows two.
2. Queue slots. Instead of setting an exact time for every post, define slots, for example weekdays at 10:00 and 18:00. Each new post takes the nearest free slot, and the schedule builds itself.
3. The link to the agent. Upload-Post has an open-source MCP server with 40 tools that works with Claude and Claude Code. Through it, the agent from step two puts the approved post into the queue itself.

Step 4. Instagram DM auto-reply triggered by a comment keyword

The last link is the one followers notice most. Someone comments a word under a reel, say "guide", and the promised material shows up in their DMs. We use ChatPlace for this: you build the flow in its visual builder, no code.

The ChatPlace flow looks like this:

1. Trigger: a keyword in the comments under a reel.
2. Greeting with a button, for example "Get it".
3. Follow check: the service makes a live request to Instagram to see whether the person follows you. Followers move on; everyone else gets an invitation to follow and comment the word again.
4. Link to the material.
5. A minute later, a second message with a link to the Telegram channel.

This is where we hit a rule that videos about auto-funnels usually leave out. Under Meta's rules, a bot can reply to a comment in DMs with exactly one message and only within 7 days. Anything after that is allowed only if the person replies, and within 24 hours of that reply. That's why the first card always has a button: a tap counts as a reply and opens the window in which the bot can send both the material and, a minute later, the channel link. Remove the button to keep the message "clean", and the flow stops at the greeting.

If you need DMs for sales rather than for handing out guides, that's the next level: qualification, objections, booking a call. We covered it in our guide on an AI sales manager in messengers.

What it costs and where "free" ends

On free plans the whole setup costs nothing. That's true, but every free plan has its own ceiling, and as you grow you hit them one by one.

The main caveat: it only comes out free if you already pay for Claude. Then the morning agent adds nothing to the bill except usage against your limits. The second bottleneck is Upload-Post: ten uploads a month is enough to try the setup, not to post daily. ChatPlace's free plan covers 200 active contacts a month; check with the service before launch what happens to replies above that limit. For a reel that unexpectedly takes off, you want to know the limit in advance.

Even on paid plans the setup comes to about $44 a month plus a Claude subscription. That's many times less than the salary from the start of the article, but it also does less work: the setup takes over the production line, not the whole job.

What stays with a person

Our opinion, which you're free to dispute: this setup replaces an SMM manager on a small account, where all of SMM is the production line, and doesn't replace one on a large account, where the main work is decisions.

What AI doesn't do here:

• strategy. The agent picks topics by your own rules. Why the account exists and where it leads followers is your call;
• brand voice. The model imitates samples; it doesn't hold a voice on its own. Without regular editing, within a month it drifts toward the internet average;
• the call on what to publish. The approval button stays with a person, and we wouldn't remove it even with perfect copy: one bad post in automatic mode costs more than a minute spent checking.

We'll be wrong for accounts where content is a stream of near-identical cards: a catalog, a timetable, promotions. There, approving every post really can be replaced with spot checks.

Plans and versions as of 3 October 2026; check the current ones with each service: they change, and routines in Claude Code are still in research preview.