Picked from the software, topic, and level of “Agentic AI Without Coding Class #9: Practice: Prompts for Social Media Captions and Product Descriptions”.
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
Hi, and welcome to the hands-on episode of Session 2. We now have everything we need: the anatomy of a prompt, templates with variables, and tidy output formats. Now let's put it all to work on the writing business owners repeat most often: social media captions and product descriptions. Our example is Sunset Coffee, and you can apply it straight away to your own products.
Step 1: Building the Sunset Coffee brand style block
The first step in this hands-on episode is finding your brand style. If you've never written it down, don't invent it from scratch. Start from evidence: five old captions that got the most likes from followers. Laras pastes those five captions into ChatGPT, then asks for a style analysis on five aspects, in table format. This uses the format technique from Episode 8. The result is a table of style patterns: a warm and casual personality, a friendly way of addressing readers, short sentences, one or two brown-toned emoji, and claims based on ingredient facts, like real palm sugar. An analysis like this is raw material. Laras still reads it, crosses out what doesn't fit, and then adds what to avoid. The end result is the following style guide. Before writing templates, we need one thing: a brand style guide. Laras writes it on five aspects. First, personality: warm and casual like a coffee buddy, not stiff or preachy. Second, how to address customers. Sunset Coffee talks to customers as you, and uses first names. Avoid dear customer or sir and madam, which feel too formal, and also guys, which feels too slangy for some customers. Third, sentences. Short, one idea per sentence. Avoid long paragraphs and all caps, which on social media reads like shouting. Fourth, emoji. Just one or two, and the choices are distinctive: a coffee cup, a brown heart, and a sunset. More than three emoji in a row makes a caption look cheap. Fifth, claims. Use product facts like real palm sugar or sugar-free. Avoid claims like best in Bandung, which are hard to prove, and never promise health benefits. We condense that guide into one block that can be pasted into every content prompt. The first part, context: Sunset Coffee, a home-based coffee shop in Bandung, selling through Instagram and WhatsApp, with customers who are students and young workers. The second part, style. This label isn't one of the five main parts, and that's fine. You're free to make your own labels as long as the content is clear: warm, casual, talking to readers as you, short sentences, and one or two emoji. The third part, product data and official prices, including the delivery rule: ten thousand rupiah in town, and free for orders over two hundred thousand rupiah. The fourth part, limits: no all caps, no best claims or health benefits, and don't make up promos. Keep this block at the very top of your Prompt Library document. Whenever prices or products change, just update them in this one place, and all your content templates use the latest data.
Step 2: A social media caption template
Now the first template: social media captions. The role block is simple: Sunset Coffee's Instagram copywriter. The context block just holds one instruction: paste the brand style block. So the template stays short, but carries the whole style guide behind it. The task block has three variables: the number of variations, the product, and the content goal. The goal can be promo, education, or story, because a promo caption and an educational caption obviously sound different. The data block has two variables: key info about the promo, and the visual being used. When it knows the visual, the AI can write a caption that fits the photo or video. The format block is the recipe for variation. Each variation uses a different approach: story, question, and straight promo. The first sentence is eight words at most, and the total length is capped by the max characters variable. And the limits block: the hashtag Sunset Coffee is required, and don't make up promos, prices, or claims outside the data. Why limit the first sentence to eight words? Because in many social media apps, long captions get cut off and only the first line shows. The first sentence has to be strong enough to make people stop scrolling. Let's run it. Dimas pastes the template along with the brand style block, then fills in the variables: three variations, the product is a five hundred milliliter cold brew, the goal is promo, buy four bottles for two hundred twenty thousand with free delivery, the visual is a reel of a dewy bottle on a rainy day, and one hundred twenty characters at most. Look at the promo info. Four bottles times fifty-five thousand is two hundred twenty thousand rupiah, which is already above the free delivery threshold. So the promo isn't made up; it's the existing delivery rule, repackaged. The result is three captions with three approaches. The first tells a story about the rainy mood, matching the visual. The second invites comments with a question. The third states the promo directly. All of them use the official price, include the Sunset Coffee hashtag, and don't overdo the emoji. The brand style comes through nicely. Then Dimas asks: are they all under one hundred twenty characters? The AI answers confidently: yes, they all are. But is that true? AI models process text in chunks of words, not letter by letter, so they often miscount characters. Let's check. Paste the three captions into Google Sheets, one caption per cell. In the next column, use the LEN function to count the characters. The story caption: eighty-nine characters. OK. The question caption: ninety characters. Also OK. But the promo caption turns out to be one hundred thirty characters, over the limit. The AI's claim was wrong. The difference is small, but when the limit is strict, such as for ads or text on an image, this mistake can get your text cut off. The lesson: for character length, numbers, and calculations, don't trust the AI's claims. Let the spreadsheet do the counting. A length test sheet like this only needs to be made once and can be used every week.
Step 3: Making variations and picking the best
Now let's fix it. In the same conversation, Dimas makes a specific request: the promo caption turned out to be one hundred thirty characters, so shorten it to one hundred twenty at most, but keep the info about four bottles and free delivery. A good request names three things: which variation to fix, what the problem is, and what must be kept. The result is a short version, eighty-seven characters according to LEN, with the key info intact. Then Dimas asks for versions in another format: two on-screen text options for the Reel, six words at most, with no hashtags. The AI gives two short options that match the caption: rain's done, cold brew's ready. And fridge stocked for the week. From one idea, we get a feed caption, a short version, and on-screen text. Why do we always ask for three approaches? Because each one has its own use. The story approach, like rain stops, the bottle sweats, builds mood and closeness with the brand. The question approach, like team iced coffee, where you at, invites followers to comment. Comments make your conversation with customers livelier. And the straight promo approach fits certain moments best: payday, a limited promo, or when there's too much stock. With three options, the business owner just picks the one that fits the moment. The last step before going live is approval. On Laras's phone, Dimas reports that three caption options are ready, and that their length and prices have been checked. Dimas also sends one of the options, the question caption, so Laras can read it right away. Laras chooses the question caption for tomorrow's feed, and asks to save the four-bottle promo caption for Friday, which falls on payday. No variation goes to waste. Dimas schedules the post and notes it in the content calendar we made in Episode 8. The working pattern is clear: the admin and the AI prepare the options, and the business owner decides.
Step 4: A product description template that doesn't invent
The second template is the product description, for a WhatsApp catalog or a marketplace. The role block: a product description writer. The context block, just like the caption template, simply pastes in the brand style block. This time the data block comes before the task, because a product description depends entirely on data. There are five variables: product name, price, size, ingredients, and storage. The task block: write a description for a certain platform. The platform is a variable, because a description for a WhatsApp catalog is shorter than one for a marketplace. The format block: one opening sentence, four key benefit points, one line on storage, and one call to order, with an adjustable character limit. And the most important block for product descriptions, limits: only use the data above. If data is missing, write a need data marker, and don't make anything up. A description that makes things up can lead to complaints, or even accusations of fraud. We test it in Claude with the cold brew product. The variables are filled in completely: a price of fifty-five thousand, a size of five hundred milliliters, West Java arabica cold brewed for eighteen hours with no sugar, storage instructions, the WhatsApp catalog as the platform, and three hundred fifty characters at most. The result follows the format: a casual opening sentence, four key benefit points, one line on storage, and a call to order. All of the content comes from the data we provided. Now for the real test. Dimas uses the same template for the gift hamper, but deliberately leaves the size, ingredients, and storage empty. The AI doesn't make up what's in the hamper. Wherever the data is empty, it writes a need data marker. This marker is easy to search for and stands out to anyone checking the text. Our limits are working. Compare that with a prompt without limits. The AI tends to fill gaps with something that sounds reasonable, such as coffee beans, chocolate, and a pretty mug. It sounds nice, but it may not be true. The checked description then goes into the catalog. On the add product page, fill in the full product name, including the size. Enter the price exactly as it appears on the official price list. Don't copy the price from the AI's text; copy it from your own data. Then paste the description. For fields that don't support bullet points, simply merge the points into regular sentences. Every online store and catalog has its own layout and character limits, and the rules can change. That's why we made the character limit a variable, not a fixed number in the template.
Step 5: One master description, many variants
Every place your content appears needs a different length. Laras made a set of house guidelines. For a feed caption, one hundred to two hundred twenty characters, with the main message in the first sentence. For on-screen text in a Reel or Story, six words at most. People only glance at it, so one idea is enough. For the WhatsApp catalog, two hundred fifty to four hundred characters, covering the price, size, and storage. And for marketplaces, six hundred to a thousand characters with bullet points, because shoppers there like to compare details. Remember, these are Sunset Coffee's internal guidelines, not official platform rules. Official limits can change, so check them regularly. The most efficient way: write one master description that's the most complete, then derive from it. Claude is asked to create three variants of the cold brew catalog description: a feed caption, Reel on-screen text, and a product title for a marketplace. Notice the last sentence: don't add any new information. The result is three variants that all come from the same data. Shortening is much safer than lengthening, because the AI only has to choose information, not add it. The marketplace product title also includes the words shoppers usually search for: cold brew, five hundred milliliters, arabica, and no sugar. As always, the length of each variant is still checked with LEN before use. All the variants are then checked in the same length test sheet, on a new tab. The feed caption is eighty-three characters, a little under the guideline of one hundred. For a caption, shorter is still fine, so Laras leaves it. The Reel text is five words, which is fine. For on-screen text, what counts is the number of words, so just read it. The marketplace title is forty-four characters, safely under the limit of sixty. And the WhatsApp catalog description is two hundred eighty-seven characters, within the guideline. All the variants are ready to use. The final decision still rests with a human; the sheet just gives honest numbers.
Step 6: Checks before posting and saving to the library
Before content goes live, run this checklist. First, prices and sizes match the official list. This is the most costly mistake, and also the easiest to prevent. Second, the length has been checked with the LEN function, not with the AI's claim. Third, no best claims or health benefits slipped through. Fourth, the style matches the brand style block: talking to readers as you, short sentences, just enough emoji. Fifth, no need data markers are left. Just use the find feature, so nothing gets missed. Sixth, customer photos and stories are only used with their permission. And seventh, the business owner has approved. For content read by hundreds of people, human approval is a required step. Now it's your turn. First, write a brand style block for your own business: personality, how you address customers, sentences, emoji, and claims to avoid. Second, use the caption template for one of your products. Ask for three variations with different approaches, then check their length with LEN. Third, write a product description, then test it once more with some data deliberately left empty. Make sure the AI writes need data markers instead of making things up. Fourth, save both templates to your Prompt Library as templates number one and two. In the next episode we complete them into ten ready-to-use templates. One more tip: under each template, note which variation got the most likes or sold the best. This simple data makes your templates sharper from month to month.
Done!
Let's recap. Consistent content comes from a brand style block pasted into every prompt, complete product data, and a format that states the approach and length. Character counts are checked with LEN, and the business owner chooses which content goes live. In the last episode of Session 2, we write prompts for replying to customers, from complaints and price questions to late orders, then gather everything into a library of ten ready-to-use prompts. See you in Episode 10.
Practice exercises
Try these on your own after watching to lock in what you learned.
Write a brand style block for your business: personality, how you address customers, sentence length, emoji, and claims to avoid.
Use the caption template for one of your products, ask for 3 variations with different approaches, then check their length with the LEN function in Google Sheets.
Use the product description template for a product whose data is incomplete, and see whether the AI writes a [NEED DATA] marker instead of making things up.
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 classes1,296 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.
