Picked from the software, topic, and level of “Agentic AI Without Coding Class #45: Practice: A Weekly Marketing Content Pipeline with Owner Approval”.
At most one email a week. Unfollow anytime with one click.
Transcript
The full narration of this video. The current sentence is highlighted while the video plays; click any sentence to jump there.
Show transcript
Introduction
Welcome to the last hands-on episode of Session 9. Today we tie everything we've learned into one system you'll actually use every week. We build a weekly marketing content pipeline for Sunset Coffee in n8n: a Monday morning schedule, a researcher, a writer, a reviewer, Laras's approval, then saving to the content calendar. At the end, we put together the full Session 9 security checklist.
Step 1: Design the pipeline and prepare Google Sheets
Before opening n8n, let's design the flow first. Every Monday at seven in the morning, a schedule triggers the workflow automatically, without anyone pressing a button. The researcher agent finds five content ideas from this week's trends and the Sunset Coffee promo calendar. The writer agent turns those ideas into five Instagram captions, complete with visual ideas and hashtags. The reviewer agent scores each caption with the four-criteria rubric from Episode 43, and asks for revisions when needed. Finally, the result goes to Laras for approval, and it's only saved to the content calendar once she approves. On Monday morning, Laras just reviews and taps a button over her coffee. The agent team does the rest. Now let's prepare the data. Create one Google Sheets file called Sunset Coffee Content Calendar, with four tabs at the bottom. The first tab is the Promo Calendar. For example, Monday, November second, is the start of the month, when many offices pay salaries. Then the rainy season, which is starting to bring frequent rain, and the weekend, a good fit for cold brew as an outing companion. There's also Heroes' Day, a national day of remembrance in Indonesia, with an important note: a greeting only, no discount. Notes like this stop the agent from making up promos. And the hampers pre-order plan starting November fifteenth. Laras only needs to fill in this tab once a month. It's the researcher's main material, so the agent doesn't have to guess the important moments for your business. The second tab is the Price List. It holds every product with its official price: the palm sugar iced coffee cup at Rp18,000, the one-liter at Rp85,000, and cold brew at Rp55,000. Then the hampers package at Rp150,000 with a no-sugar version, and the delivery rule: Rp10,000 in town, free over Rp200,000. This is the only source of prices for the writer and the reviewer. The other two tabs, Content Calendar and Log, stay empty for now. The workflow will fill them in. Just add the column headers: date, product, caption, visual idea, score, and status. You only prepare this data once, in about fifteen minutes.
Step 2: Monday morning schedule and the researcher agent
Open n8n and create a new workflow called Weekly Content Pipeline. The first node is a schedule trigger. Click that node, and its settings panel opens on the right. Set it to repeat every week, on Monday, at seven in the morning. Make sure the time zone is Asia Jakarta, or your own time zone, so the workflow doesn't run at the wrong hour by following the server's time zone. The second node reads Google Sheets. Choose the Sunset Coffee Content Calendar document, the Promo Calendar tab, then filter for just the next seven days. The researcher doesn't need to read the whole month's plan. The third node is the researcher AI Agent. Paste the researcher instructions from Episode 41. Its incoming message, shown in yellow, takes the list of moments from the calendar node. Attach an AI model as the researcher's brain. For finding ideas, a low-cost model is usually enough, following the cost-saving tips from Episode 44. Then give it one tool: web search, so the researcher can see this week's trends, such as the weather or topics people are talking about. The researcher only has read tools, nothing that sends anything. Click test step. The three main nodes get green check marks. Menu names in your n8n may differ slightly from this simulation, but the setup is the same. Let's look at the researcher's work. Click the researcher node to see its output. First, the researcher reads the promo calendar and finds four moments this week. Second, it uses the search tool to check the Bandung weather and warm drink trends. It turns out this week is expected to be rainy, information that isn't in the promo calendar. The result is a table of five ideas, one for each weekday: an office coffee stock-up at the start of the month, warm coffee for a rainy day, the real palm sugar story, the hampers pre-order, and cold brew for the weekend. Notice that the researcher doesn't write captions or mention prices, exactly as its limits say. Tuesday's idea came from the weather trend, not from the calendar. That's what the search tool is for.
Step 3: The writer agent: 5 captions and visual ideas
We continue from the researcher node, which already has its check mark. Before the writer, insert a Google Sheets node called get prices, which reads the whole Price List tab. Why a separate node? Because the writer must never remember prices on its own. If prices go up next month, Laras only changes one cell, and the next week's captions use the new price automatically. Then add the writer AI Agent. Paste the writer instructions from Episode 41: a warm style, greeting customers warmly, two emojis at most, and prices only from the list. Its incoming message combines two things: the ideas from the researcher and the price list from the previous node. Both are pulled in by dragging fields from the data panel, with no code to type. Attach an AI model. For writing, pick the model whose language reads best. The writer gets no tools at all, because its only job is to write. Click test step. Get prices and the writer get green check marks. Now let's read the output. The Monday caption: start of the month, time to stock the office with coffee. The one-liter price of Rp85,000 matches the list, there's one emoji, plus a visual idea of a bottle on an office desk and three hashtags. The Tuesday caption uses the researcher's rain idea: palm sugar iced coffee for Rp18,000, order on WhatsApp and we'll deliver. The visual idea is a cup by a fogged-up window. Now look at the Thursday caption. Year-end hampers, from Rp125,000. But in the Price List, the hampers is Rp150,000. The writer can still make mistakes, even when it's given the data. With two more captions for Wednesday and Friday, that makes five in total. We're leaving Thursday's price mistake in on purpose, because in the next step we'll see whether the reviewer catches it.
Step 4: The reviewer agent with a rubric
After the writer, add the reviewer AI Agent. Its instructions contain the rubric from Episode 43: accurate, on-brand, complete, and safe, scored from one to five. Its incoming message contains all the captions and the same price list the writer used. The reviewer must compare every number against this list, not against its memory. Attach an AI model. A tip: if you can, use a different model from the writer's, so they don't think in exactly the same way and mistakes are easier to catch. The next node is an IF, which asks one question: did every caption pass? If yes, the flow moves on to the step for Laras's approval. If something needs revising, that caption goes back to the writer once more with the reviewer's notes, then gets checked again. We limit it to one revision, so costs don't balloon. If it still fails, the caption still goes to Laras with a warning flag. Let's test it. Every node gets a check mark, including the revision path. That means the reviewer found something. Let's look at its report. The reviewer sends a score table for all five captions. Four pass with scores of four or five. The Thursday caption gets an accuracy score of one, and its status is revise. The notes are clear: the hampers price says from Rp125,000, but the Price List says Rp150,000. The reviewer also suggests adding the pre-order info starting November fifteenth, to make it complete. The Thursday caption goes back to the writer with those notes. This is the first and only revision, matching the limit we set in the IF node. After the revision, the Thursday caption scores five on every criterion and passes. The price mistake was caught before it ever reached Laras, let alone a customer.
Step 5: Send to Laras for approval
The next step is approval. After the IF node says everything passed, add a build summary node. It combines the five captions, visual ideas, and scores into one numbered message, so Laras can read it in one go. Then the send and wait for approval node, with exactly the same pattern as Episode 42. The channel is Telegram, and the recipient is Laras. There are two buttons: approve all, or request changes. The time limit is Tuesday at twelve noon. If there's no reply, the workflow stops, and no content gets marked as ready to publish. After Laras replies, a second IF node checks her choice: approve all, or request changes. If she approves, the captions are saved to the Content Calendar tab with the status approved, along with their posting dates. If Laras requests changes, her notes are saved in the Log tab, and the caption she named is marked as needing revision. The other captions she didn't mention are still saved as approved. This is what Laras receives at seven oh four. One message with a summary of five posts for November second to sixth, with an average score of four point eight. The hampers is now listed correctly, at Rp150,000, with the pre-order info. Below it are two buttons: approve all, or request changes. Laras doesn't need to open her laptop, n8n, or Google Sheets at all. Laras reads it while getting the shop ready. All the prices are right, but she wants one photo swapped. She taps request changes at seven twenty. She writes a short note: for number three, use the photo of the palm sugar farmer we took yesterday. The rest is fine. More authentic visual ideas often come from the business owner herself. The bot confirms: number three is marked for revision, and the other four captions are saved. Laras spent about fifteen minutes in total, mostly reading calmly. If Laras is busy and doesn't get to her phone until the afternoon, that's fine. The workflow keeps waiting until the Tuesday noon time limit, without sending or saving anything.
Step 6: Save to the content calendar and run a full test
Open the Content Calendar tab. Five new rows appear automatically, in green. Each row has a date, product, caption, visual idea, score, and status. The Wednesday row has the status revise from Laras, with the visual idea already changed to a photo of the palm sugar farmer. The full note is in the Log tab. Thursday has a perfect score after the reviewer's revision, and Friday is ready for the weekend. Every price matches the Price List. From here, whoever does the posting, such as Nadia or Dimas, just copies the captions from the calendar. No more hunting for ideas at the last minute. Now let's look at the full workflow from left to right. This is what we built today. The Monday seven o'clock schedule triggers the researcher. The researcher uses the trend search tool and reads the promo calendar that Laras fills in once a month. The writer writes five captions with the official price list from Google Sheets, not from memory. The reviewer scores them with the rubric and asks for a revision when needed, once at most. Then comes Laras's approval on Telegram, with a time limit of Tuesday noon. Approved results are saved to the Content Calendar, and revision notes to the Log tab. To keep it compact, the IF and build summary nodes aren't shown in this picture. Finally, flip the active toggle at the top. Starting next Monday, the pipeline runs on its own. The full test run took about two minutes, plus sixteen minutes of waiting for Laras.
Step 7: Session 9 output: content system + security checklist
Let's look at the impact. Before, finding ideas and writing five captions by hand could take about four hours every week, often done late at night after the shop closed. Now, Laras only needs about twenty minutes to review and approve on Telegram. And most important, no caption with a wrong price reaches customers, because the reviewer and Laras's approval act as two layers of filtering. These numbers are illustrations. Track your own time, because we'll use it to measure impact in Session 10. Now the second part of the Session 9 output: the full security checklist. Use it for all your agents and workflows, not just the content pipeline. First, as little customer data as possible, with no ID cards, bank account numbers, or card details. Second, the data usage setting for model training has been checked in every app, along with the date it was checked. Third, agents only have the tools they truly need. The researcher only reads, the writer has no tools, and no agent holds payment access. Fourth, customer messages and web page content are treated as data, not commands. This is the defense against prompt injection. Fifth, sending, paying, and changing prices require human approval, with amount limits and a time limit. Sixth, the agent answers only from data, and dares to say I don't know when the data isn't there. Seventh, test questions and the rubric are run regularly, and every approval decision is saved in the log. Eighth, monthly costs are monitored and compared with the work hours saved. If all eight boxes are checked, your system isn't just smart, it's also safe and worth keeping. Now it's your turn. First, replace the sample data with your own business data: the promo calendar, the price list, and your brand style guide. If you're a teacher like Mr. Hendra, swap the captions for weekly lessons or practice questions. Second, adjust the schedule and the approval channel. Sunday evening might suit you better, or you might check email more often than Telegram. Third, run it for two weeks under close supervision. Read every result before approving, and note the mistakes that come up so you can improve the instructions. Fourth, save this workflow and the security checklist. Both become the main material for your final project in Session 10.
Done!
Congratulations, Session 9 is complete. You now have an automatic content system run by an agent team, checked with a rubric, and only saved once the owner approves, along with a security checklist for all your agents. In the next episode we start Session 10, the final session: building an automation map of your business as the foundation for your final project. Have your list of five tasks from Session 1 ready, along with all the workflows you've built. See you in Episode 46.
Practice exercises
Try these on your own after watching to lock in what you learned.
Build the same pipeline for your business: replace the promo calendar, price list, and style guide with your own.
Run the pipeline for two weeks in a row under close supervision, then note how much time you spend reviewing and how many captions needed revisions.
Fill in the Session 9 security checklist for all your agents and workflows. Save it with the workflow as material for your final project in Session 10.
Ask the class material
Type a question or keyword; we find the part of this episode that explains it and jump the video there.
My notes
Notes are saved with the current video time and are only visible to you.
Questions & answers
Stuck on a step? Ask here; the instructor and other learners can answer.
No questions yet.
Your instructor
Cerdas Inovasi Digital Team
Software developers, BIM & 3D engineers, and AI practitioners
53 video classes1,286 episodes
We build software, BIM models, games and AI automations for clients every day, and turn that hands-on work into step-by-step classes: every step is shown on screen and comes with a PDF guide you can keep.
