ChatGPT prompts for a content calendar, tested on 936 weekdays

We expected the dates to be the problem. A content calendar is a list of dates with weekdays next to them, and Christmas Day and New Year's Day both fall on a Friday this season, so we wrote a script to check every date three chat models gave us. We ran five prompts, 31 times in all, on ChatGPT, Gemini and Claude, for an invented bakery planning its posts from 1 December 2026 to 10 January 2027.

We were wrong about the dates. All 936 dates the three models wrote next to a weekday were correct, and every run with a brief posted on exactly the 17 days we asked for. What went wrong was the business. ChatGPT twice told customers the bakery was "open Tuesday to Sunday, 7am-3pm" in Christmas week, when it is shut on the 25th and 26th. Gemini named sourdough, baguettes and focaccia we never mentioned and invented "order forms" behind "the link in our bio". One prompt, prompt 5 below, fixed both in all six runs we gave it.

Can ChatGPT make a social media content calendar?

Yes, and on dates it is reliable: none of the three models we tested put a wrong weekday next to a date in 31 runs. The risk is in the facts about your business, which a model fills in when you leave a gap and sometimes misapplies when you do not. This is what we counted.

The content calendar prompt that worked

Start here. This is prompt 5, written after we had read the 25 runs of prompts 1 to 4 and seen what went wrong. It gives the posting days, the usual hours and every exception week by week, says exactly how pre-orders are taken, and tells the model what not to add. Swap in your own business and keep the structure.

Make a social media content calendar for Wrenfield Bakery, a small independent bakery, from Tuesday 1 December 2026 to Sunday 10 January 2027. We post on Instagram and Facebook on Mondays, Wednesdays and Fridays. Opening hours: Tuesday to Sunday 7am to 3pm, closed Mondays, with these exceptions. Christmas week: open Tuesday 22, Wednesday 23 and Thursday 24 December 7am to 3pm, closed Friday 25 and Saturday 26 December, open Sunday 27 December 7am to 3pm. New Year week: open Tuesday 29 and Wednesday 30 December 7am to 3pm, Thursday 31 December 7am to 1pm, closed Friday 1 January, open as usual from Saturday 2 January. Christmas cake and mince pie pre-orders are taken in the shop and close on Saturday 19 December. Rules: when a caption mentions opening hours, give only the hours for the days it talks about, using the exceptions above, never the usual weekly hours for a week that has an exception. Name no products except Christmas cake and mince pies. Do not invent prices, offers, ordering methods, statistics, awareness days or national days. Give me a table with these columns: Date, Day, Post idea, Caption (one or two sentences). After the table, list anything you had to assume.

We ran it twice on each model. All six calendars used the 17 posting days, and every caption that gave hours in Christmas or New Year week gave the right ones. ChatGPT's Monday 21 December caption read: "We're closed today and will be open Tuesday 22, Wednesday 23 and Thursday 24 December, 7am-3pm." No run named a product beyond Christmas cake and mince pies, or a way to order other than the shop. And all six ended with a list of assumptions, which is where the useful questions were: Gemini assumed customers would collect pre-orders in Christmas week, and Claude assumed pre-orders opened on 2 December. Both are things to confirm, and one Gemini run already wrote its guess into a caption as fact: "We are delighted to see everyone picking up their Christmas cake and mince pie pre-orders today".

The line that did the work is the one about weekly hours. It is the mistake the earlier prompts made, as the next sections show, and none of the plainer rules we tried, "use only the facts above" included, prevented it.

What happens with "make me a content calendar for December"?

Make me a social media content calendar for December.

You get a calendar for nobody. All seven runs wrote for a generic brand, with posts such as "Monday motivation" and "Our most popular product/service", and six of them planned a post for every one of the 31 days of December; Gemini's second run planned 24. If you post twice a week, that is a calendar you will abandon by the 8th. How often a small business can realistically post is its own question, and we have answered it here.

All seven planned December 2026: every run's weekdays fitted 2026, and every run that named the year of December named 2026. Each app, and our system prompt, told the model today's date, so we did not test what happens without it. ChatGPT's single run gave dates and post ideas with no weekdays, then ended with a questionnaire asking what the business does and how often it posts. Those are the questions the prompt should have answered in the first place.

Claude and Gemini filled the month with awareness days. Most were on the right dates for 2026: GivingTuesday on Tuesday 1 December, which is the date GivingTuesday itself gives; International Volunteer Day on 5 December, a United Nations observance; and Hanukkah beginning on Friday 4 December in all three of Claude's runs and in Gemini's first, which matches Hebcal's calendar, where the first candle is lit that evening. Gemini's third run put "Hanukkah Begins" on Sunday 6 December, two days late. Several others, such as National Cookie Day and Ugly Sweater Day, are commercial "national days" with no official body behind them; whether they belong on a bakery's feed is a choice, not a fact a model should make for you.

Did the models get the weekdays right?

Yes, every time. We gave prompt 2, a full brief, three times to Gemini and Claude and twice to ChatGPT, then gave prompt 3, which adds an instruction to check every weekday against a calendar, and prompt 4, which lists the 17 posting dates outright. Across all five prompts the script found 936 dates written next to a weekday, from table rows such as "25 December 2026 | Friday" to sentences such as "closed on Friday 25 December and Saturday 26 December". Not one was wrong, and every brief run posted on the 17 Mondays, Wednesdays and Fridays and on no other day.

So telling a model to check its dates, prompt 3, made no measurable difference, because there was nothing to fix. It did change the answers in one visible way: all three models wrote out the Mondays, Wednesdays and Fridays before the table, which makes the calendar easier for you to check. Giving the dates outright, prompt 4, is the safest habit when a date really matters, but on this run it was not needed.

We say this plainly because the advice we expected to write was that chat models get weekdays wrong. On these models, with today's date known, they did not. If you use an older or smaller model, or a tool that does not pass today's date, check anyway; the checklist at the end takes two minutes.

What did the models get wrong in a content calendar?

Opening hours in the weeks that had exceptions. This is prompt 2, the brief most people would write:

Make a social media content calendar for Wrenfield Bakery, a small independent bakery, from Tuesday 1 December 2026 to Sunday 10 January 2027. We post on Instagram and Facebook three times a week, on Monday, Wednesday and Friday. Facts you can use: we are open Tuesday to Sunday, 7am to 3pm, and closed on Mondays. Christmas cake and mince pie pre-orders close on Saturday 19 December. We are closed on 25 and 26 December and on 1 January. On 31 December we open 7am to 1pm. We reopen as usual on Saturday 2 January. Give me a table with these columns: Date, Day, Post idea, Caption (one or two sentences).

Every fact a customer needs is there, and most captions used them well. The holiday closures were stated correctly in every run, and so was the short New Year's Eve. The trouble was the Monday posts at the start of Christmas week and New Year week, where a model wanted to remind people when to come in and reached for the usual hours. ChatGPT, Monday 21 December, in the run of prompt 4:

Christmas week is here! We're open Tuesday to Sunday, 7am-3pm, so pop in and see us before our Christmas closure.

And Monday 28 December in the same run: "We're back open Tuesday to Sunday, 7am-3pm, so come and visit us this week." That week the bakery closes at 1pm on Thursday and does not open on Friday. ChatGPT made the same two mistakes, in nearly the same words, in its run of prompt 3, and Claude wrote the 28 December version once, in its third run of prompt 2. Gemini never did: in every run of prompts 2 to 4 its Christmas week caption said the bakery was open Tuesday to Thursday, in those or nearly those words, which leaves out the Sunday opening but promises nothing false.

Notice which ChatGPT runs made the mistake: the two that told it to use only the facts we gave. The usual hours are one of those facts, so a rule against inventing does not catch a true fact used in the wrong week. That is why prompt 5 spells out each holiday week and adds the rule about weekly hours.

Do AI models invent products for a content calendar?

Gemini did, every time it was not told not to. All three of its runs of prompt 2 named breads and bakes our bakery never mentioned: sourdough and baguettes in all three, croissants in two, and in one run "warm rosemary focaccia and toasted sandwiches". All three also invented a way to order: "Order forms are available via the link in our bio" in the first, "send us a message" in the second and "order in-store or online" in the third. Its third described the bread as "made from just four natural ingredients". ChatGPT and Claude kept to generic words such as "loaf", "pastries" and "coffee". Where they said how to order, they said in the shop, which our prompt did not say but which was a safe guess, and two Claude runs said instead that they did not know how we take orders and suggested we add it.

For a real bakery each of these is a small promise. A customer who reads "focaccia" on Instagram and finds none on the shelf does not blame the model. With the rule "do not invent statistics, prices, products or opening hours", in prompts 3 and 4, no run of any model named a product, an awareness day or a statistic, and none mentioned ordering online. Several still told customers to "pop in" to order, and one Claude run added "or get in touch with us"; prompt 5 settles that by saying where pre-orders are taken.

Which AI is best for a content calendar?

On this test, Claude, for one reason that matters more than the captions: it tells you what it guessed. In 6 of its 7 brief runs it added a note like this one, unprompted:

You didn't mention any change for Thursday 24 December, so I assumed normal hours (7am to 3pm). Please confirm that before posting on 23 December.

That is exactly the question a bakery owner should answer, and neither ChatGPT nor Gemini listed any assumptions until prompt 5 asked them to. Gemini was the most reliable on hours and the most inventive about products. Claude wrote the shortest captions; ChatGPT made the most hours mistakes. Seven to twelve runs a model is enough to show a tendency, not a ranking, so the practical answer is the same for all three: use prompt 5, then check the result.

If you want captions as well as dates, the same habits apply to the prompts in ChatGPT prompts for social media, Gemini prompts for social media and Claude prompts for Instagram posts.

How do you check an AI content calendar before you schedule it?

Read only the captions that make a promise. In our runs the errors were never in the post ideas or the tone; they were in sentences that tell a customer when to come, what to buy or how to order. Those are a handful per month, and they are the ones to read against your own diary.

For what to post in December in the first place, with the 2026 dates, see Christmas social media posts for small business. For the New Year images themselves, Gemini prompts for New Year 2027 posters has seven tested prompts.