Content plan
A content plan is a table: date, topic, platform, format, status. The agent fills it with suggestions, and you keep the topics the business needs and strike out the rest.
The agent takes topics from your sources: customer questions from chats, common objections from the sales team, industry news, past posts with a good response. The more precise the sources, the fewer generic topics like 5 tips for business.
An easy start: once a week the agent suggests 10 topics, and the editor keeps 3. After a month you see which sources give the topics that get chosen, and you adjust the agent’s instruction to them.
Anthropic advises starting simple and adding complexity only when needed. A plan for a month often needs no agent: it is one task for the model once a month with a ready instruction. You need an agent when topics must change every week based on new data.
This setup is built in n8n, an automation builder from Berlin. Its license is fair-code (Sustainable Use License), not an OSI open license: you may install it on your own server and change it, but you may not resell it as a cloud service.
Drafts in your style
Style is set with examples, not adjectives. The model will read a request to write in a lively and expert way in its own way, while five of your good posts show it the length, tone and structure.
The Claude documentation calls examples one of the reliable ways to set the format, tone and structure of an answer and advises giving 3–5 examples. The examples should be similar to your task and at the same time differ from each other.
The agent’s instruction, that is, the system prompt with standing rules for the model, contains: who the reader is, what not to write about, the company glossary and 3–5 sample posts. Facts go in a separate list: prices, terms, addresses. The agent takes them only from there.
Each platform has its own format: a short post for Telegram, a long text for a blog, a caption for a video. Give the agent a separate example for each platform, or all texts will come out the same length and tone.
The agent orders images for posts from an image model, using the description from the draft. The editor checks them like text: the product, the logo, any words in the image.
A draft is not a finished post. Its job is to save the author time on the first version, while the meaning and the numbers stay with a person.
Publishing on schedule
The n8n setup is built from ready nodes. The steps below were checked against the n8n documentation.
1. Add the Schedule Trigger node (start on a schedule). In the Trigger Interval field, choose Days or Weeks. The schedule follows the workflow’s time zone, or the n8n server’s time zone if none is set.
2. Add the AI Agent node and connect a model and at least one tool to it, for example reading the content plan table. The agent’s instruction holds the style rules and examples.
3. Add a Telegram node with the Send and Wait for Response operation and the Approval response type. The draft goes to the editor, and the workflow pauses until the editor decides.
4. After approval, a second Telegram node with the Send Message operation publishes the post. In the Chat ID field, enter the channel in the @channelusername format, and add the bot to the channel in advance.
5. Save and publish the workflow in n8n. Without this, the Schedule Trigger node will not start it on schedule.
For the first two weeks, send approved posts to a closed test channel, not the main one. This way you see how a post looks in the feed and fix the format before customers see it.
Turnkey publishing on several platforms, with replies to comments and requests sent to the CRM, is described on the page about AI for business social media.
Editor and fact-checking
A model sometimes writes confident and wrong text. The Claude documentation calls this hallucination and warns directly: methods to reduce errors help, but they do not remove them fully, and important points must be checked.
So there is always a person in the setup. The editor approves every post until the agent has a month of posts without factual edits; after that, checks can be done selectively.
Three methods from the same documentation take some load off the editor. Allow the model to answer I do not know. Give it only your documents as a source. Ask it to back each claim with a quote and to delete what it cannot back.
The editor always checks numbers, prices, dates and names. A wrong price in a public post costs more than five minutes of checking.
The editor also checks rights: other people’s photos, quotes and logos do not go into a post without permission, even if the agent found them in open access.
Link the agent to a metric. The number of posts says nothing; count the requests and clicks from posts, or the agent will only speed up the flow of texts nobody reads.