Monday, 10 am. Three clients, three calendars, and one of them has a festival launch on Friday. A customer has commented "where is my order" on a post from last week. The monthly report is due on the 5th and the numbers are in four different exports. AI prompts for social media managers help most with the repetitive middle: first-draft calendars, caption variants, sorting the inbox, and writing the commentary on a report. They help least with anything that speaks for the brand in a crisis, or with any figure you didn't compute yourself.
These are untested templates; I haven't run them on a live account, and the examples are invented. The one piece of arithmetic here was run with a short script, and I paste the real output. For platform rules, I read one primary page this session, YouTube's help page on disclosing AI content, on 2026-10-09. I could not read Instagram's or Meta's policy pages in this session, and I could not read the text of India's advertising guidelines for influencers, so for those I tell you what to check instead of what the rule says. For scripts and hooks, see Instagram Reels script prompts; for professional posts, LinkedIn content prompts. For a business with one location, prompts for restaurant owners covers reviews and menu posts.
How do I build a content calendar with AI that isn't generic?
A calendar is a planning document, so give the model constraints and facts, not just a niche. Generic input gets generic ideas.
You are helping a social media manager plan [MONTH] for [CLIENT CODE].
BRAND FACTS (the only claims you may use):
- What they sell and to whom: [DESCRIPTION]
- Products/offers running this month, with dates: [LIST]
- Things we may NOT say (legal, regulatory, claims we cannot support):
[LIST]
- Tone guide (3 do's, 3 don'ts, two sample posts we approved): [PASTE]
- Platforms and formats we produce: [E.g. Instagram reel, carousel,
static; LinkedIn text; YouTube Short]
- Capacity: [N posts per week, who produces the creative, review lead time]
- Dates that matter for this client (festivals, launches, sale days):
[LIST, from the client's own plan or a calendar I checked]
OUTPUT:
A table: date, platform, format, topic/angle, the proof or asset we need
(for example: a photo of X, the price from the client, a customer
permission), the CTA, and the person who must approve it.
RULES:
- Max [N] posts per week. Never schedule beyond our capacity.
- Do not invent offers, prices, discounts, testimonials or product claims.
If a post needs one, write [NEED FROM CLIENT: ...].
- Do not assume festival dates. Use only the ones I listed.
- Variety: no more than two consecutive posts of the same format.
- Flag posts that depend on a person or customer appearing in the creative,
because we need written permission.
The "proof or asset we need" column is the useful part. It turns a calendar of topics into a to-do list for the client and the designer, and it makes the dependency on a missing price or photo visible on Monday instead of Thursday.
Always take festival and sale dates from your own calendar. A model can get a lunar-calendar date wrong by a day, and a "Happy Diwali" post going out on the wrong day is a very public mistake.
How do I write captions that stay on brand and inside the rules?
Give the model approved posts to imitate and a list of claims it must not make. Then ask for options, not one final.
Write 3 caption options for this post.
POST: [PLATFORM, FORMAT, WHAT THE CREATIVE SHOWS]
BRAND VOICE: [PASTE TONE GUIDE + 2 APPROVED CAPTIONS]
MUST INCLUDE: [OFFER, DATE, LINK PLACEHOLDER, HASHTAGS from the approved list]
MUST NOT INCLUDE: [CLAIMS WE CANNOT MAKE, COMPETITOR NAMES, PRICES NOT GIVEN]
For each option: caption, a one-line reason it fits the voice, and any
line that makes a claim someone needs to check, tagged [CHECK: ...].
Rules: no invented statistics, awards, reviews or customer quotes. No
"best", "No.1" or "guaranteed" unless I have written it in MUST INCLUDE.
Keep within [N] characters. If the creative needs a disclosure (paid
partnership, AI-generated realistic content), say so at the end.
Superlatives and health or financial claims are where brand accounts get into trouble. India's advertising self-regulator publishes guidelines for advertisers and influencers (its site is ascionline.in); I could not read the guideline text in this session, so I won't summarise it. Read the current version, and ask the client's legal or compliance contact about regulated categories such as health, finance and education, before you rely on a model's idea of what is allowed.
What do I check before disclosing AI use or sponsored content?
Disclosure rules differ by platform and are updated. Build a checklist from each platform's own page and date it.
What I verified: YouTube's help page "Disclosing use of GenAI content", read on 2026-10-09 via a text-extracted copy of the page. It says creators must disclose when they use AI to meaningfully alter or generate photorealistic content, and gives these categories: making a real person appear to say or do something they didn't, altering footage of a real event or place, generating a realistic scene that didn't occur, and music that's the main focus of a video. It lists as not needing disclosure things like production assistance (scripts, thumbnails, titles, outlines), captions, and cloning one's own voice for voice-overs or dubs. The disclosure is a setting under "AI use" when you upload in YouTube Studio, and the page says YouTube may also label content automatically, for example when it carries C2PA metadata, and that disclosing does not limit a video's audience or monetisation eligibility. It also says that consistently not disclosing can lead to a label being applied manually or to penalties. Read it directly, because this is my summary, not a quote, and the page can change.
What I did not verify: Instagram, Facebook, X and LinkedIn rules for AI labels and branded content, and any legal requirement in India to disclose paid partnerships. Don't take a number or a rule from a model on these. Use this prompt to produce the checklist from the actual page:
Below is the text of [PLATFORM]'s current policy page on [AI-generated
content labels / paid partnerships], copied by me on [DATE]:
[PASTE THE PAGE TEXT]
Produce a pre-publish checklist for a brand account, using ONLY this text.
For each item, quote the sentence it comes from. If the page does not say
something, write NOT STATED ON THIS PAGE. Do not add rules from memory.
Paste the page text yourself, with the date. The checklist is only as good as the page you pasted.
How do I triage comments and DMs without sending the wrong reply?
Sort first, draft second. Messages in a brand inbox range from "price?" to a complaint that should go to a senior person. Teach the model the categories and the escalation rules.
CATEGORIES AND RULES:
- A. Simple question with an answer in the FAQ below: draft a reply.
- B. Praise: draft a short thanks. Do not offer anything.
- C. Complaint about an order, refund or delivery: DO NOT reply publicly;
draft a short "we're sorry, please DM us your order number" and mark
ESCALATE: support.
- D. Health, safety, legal, defamation or a threat: mark ESCALATE: manager.
Do not draft a reply.
- E. Spam or abuse: mark HIDE/REPORT. Do not draft.
- F. Anything unclear: mark UNSURE. Do not guess.
FAQ (the only facts you may use): [PASTE]
TONE: [PASTE]
MESSAGES (personal details removed): [PASTE, numbered]
For each message give: number, category, draft or action, and the line
of the FAQ you used.
Check UNSURE and ESCALATE rows personally. A model can be fluent and polite to someone who has just described an allergic reaction, and that is the wrong reply. Never paste customers' phone numbers, addresses or order details into a chat. The PII redaction post shows how to strip them before sending.
How do I write a monthly report that holds up?
Compute the metrics yourself, then ask the model for commentary. The mistake to avoid is mixing definitions. I ran this on five invented posts, using interactions (likes, comments, saves, shares) divided by reach:
P1 reel reach 12400 interactions 586 rate by reach 4.7%
P2 carousel reach 5200 interactions 409 rate by reach 7.9%
P3 static reach 3100 interactions 86 rate by reach 2.8%
P4 reel reach 8700 interactions 283 rate by reach 3.3%
P5 carousel reach 4800 interactions 332 rate by reach 6.9%
reel posts 2 pooled rate 4.1%
carousel posts 2 pooled rate 7.4%
static posts 1 pooled rate 2.8%
These numbers are invented and are not a benchmark. Look at the reels: the two per-post rates are 4.7 and 3.3 percent, whose simple average is 4.0, while the pooled rate (total interactions over total reach) is 4.1. The gap is small here and can be large when reach varies a lot. Pick one method and state it on page one. Sample size is tiny too: five posts show a pattern worth testing, not a conclusion.
Write the commentary for a monthly social report for [CLIENT CODE].
METRICS (computed by me; definitions below; do not recalculate):
[PASTE TABLE]
DEFINITIONS: [Engagement rate = interactions / reach; interactions =
likes + comments + saves + shares; reporting window = ...]
WHAT WE DID: [CAMPAIGNS, DATES]
WHAT THE CLIENT ASKED FOR: [GOALS]
Produce:
1. Three observations that follow directly from the table. Name the
posts.
2. For each, one possible explanation labelled HYPOTHESIS, and one
cheap test for next month.
3. A section "What this data cannot tell us" (sample size, paid vs
organic mix, platform changes, anything missing).
4. Two recommendations tied to the observations.
RULES: No industry benchmarks. No claims about causes as if proven. No
sales or revenue claims unless in the table. No prediction.
The "what this data cannot tell us" section is the one a client will quietly appreciate and an agency boss will quietly cut. Keep it.
A worked example (illustrative)
An invented scenario to show the output shape, not a real run. Inputs: a home-decor brand, a Friday festive launch, and these overnight comments: "Is this available in Pune?", "My order from last week hasn't come", "Your product caused a rash on my hand!".
What good triage looks like (shape):
1. A, draft: "Thanks for asking! [FAQ: delivery pincode checker link].
Share your pincode and we will confirm." Source: FAQ line 3.
2. C, ESCALATE: support. Draft: "We're sorry about the delay. Please DM
your order number and we'll look into it." No public detail.
3. D, ESCALATE: manager. No draft. Reason: health complaint about a
product.
Item 3 is the point. The model sorted it and stepped back, and a person now writes that reply.
Failure modes to check for
| What goes wrong | Why | What to do |
|---|---|---|
| Wrong festival or sale date | Model recalls calendar dates | Use only dates you checked |
| Invented offers or claims | It fills gaps to sound persuasive | [NEED FROM CLIENT] rule; claim checks |
| Policy "facts" from memory | Platform rules change | Paste the current page; dated checklist |
| Cheerful reply to a complaint | Tone beats judgment | Category rules; ESCALATE list |
| Report numbers differ month to month | Definitions drift | Write definitions on page one |
| Causes stated as facts | Small samples, one month | HYPOTHESIS label; "cannot tell us" section |
| Customer details pasted | DMs contain phone and address | Strip first; client codes |
What not to automate
- Crisis and complaint replies. Anything involving harm, refunds, legal threats or a named person needs a human and often a manager.
- Disclosure decisions. Whether a post needs an AI or paid-partnership label is a policy check against the platform's current page, not a model's opinion.
- Claims. Health, finance and comparison claims need the client's sign-off, and compliance review in regulated categories.
- The numbers. Compute metrics and define them yourself.
- Posting itself. Schedule from a human-approved calendar. Auto-publishing model output is how a placeholder ends up live.
For the habit of limiting a model to supplied facts, see avoiding hallucinations and giving context. To turn posts into visuals, carousel and infographic prompts goes through layout and text checks, and the prompt library has a social caption template to start from.
