How to write a brief that gets you a usable AI social media draft
Most disappointing AI captions come from thin briefs. A five-part brief, worked examples and the follow-up requests that turn a first draft into a post you would publish.
AI writing assistants are now built into most social media tools, and most people’s first experience is the same: you type “post about our autumn menu”, you get something cheerful and generic, and you rewrite it yourself. The assistant is not the problem. It can only use what you give it, and “our autumn menu” gives it almost nothing. This guide covers what a useful brief contains, how to set up a brand voice once so you do not repeat yourself, and how to work with the draft afterwards.
Why short prompts produce generic posts
An AI model writing a social post has to decide many things at once: what the news is, who it is for, why they should care, what to do next, how the brand sounds and how long each network allows. If the brief does not answer those questions, the model fills the gaps with the most typical answer. Typical answers are what make a post sound like every other post.
The fix is not a cleverer prompt. It is the same thing a good human copywriter asks for: a brief.
The five parts of a good brief
A brief for a single post rarely needs more than five short lines.
- What is happening. The concrete news or point: a product, an event, a change, a tip. Include the specifics: names, dates, times, prices, places.
- Who it is for. The audience for this post, not the brand’s whole audience. “Regulars who come in before work” is better than “customers”.
- Why they should care. The one benefit or reason that matters to that audience.
- What you want them to do. Book, visit, read, reply, save. One action.
- Anything that must or must not appear. A required disclosure, a link, a hashtag, a word the client never uses, a claim legal has not approved.
A thin brief and a useful one
For a fictional café, Lantern Street Coffee:
| Thin brief | Useful brief |
|---|---|
| Post about our autumn menu. | New autumn menu starts Monday 6 October: spiced pear latte, maple oat porridge, pumpkin and sage toastie. For regulars who grab breakfast before work. Why care: porridge and toastie are ready in under five minutes. Action: try it this week. Must include: available until 30 November. Avoid: “pumpkin spice”. |
The second brief is still only five lines, and every line removes a guess. The model no longer has to invent dates, products or a reason to visit, so it will not.
Put every fact the post needs into the brief. A well-instructed assistant should not invent prices, dates or claims, but that only helps if the real ones are there to use.
Set up the brand voice once
The parts of a brief that never change (how the brand sounds, words to use and avoid, standard hashtags) belong in a brand voice, not in every brief. A useful brand voice covers:
- Who the brand is, in a sentence or two
- Who it talks to
- How it sounds: three adjectives with a “but not” each, such as friendly but not jokey
- Words to use and avoid, including product names and terms the brand never uses
- Hashtags it always uses
- Emoji: yes, no, or only in certain places
Write it once, check it with the client, and reuse it for every post. If you work across several clients, keep a separate brand voice per client; mixing them is the quickest way to get a bank’s post that sounds like a bakery.
Ask for a version per network
One caption rarely suits every network. Instagram wants a strong first line and hashtags; X wants the point in a fraction of the space; LinkedIn wants the reason it matters to a professional reader; TikTok captions support the video rather than carry the message. Ask for a version per network, and tell the assistant the networks you are posting to so it can respect their limits and link rules. Our guide on adapting one post for seven networks lists what to change for each.
Follow-up requests that improve a draft
Treat the first draft as material, not a finished post. Small, specific requests work better than “make it better”:
- Stronger opening: ask for a first line that states the benefit, not the announcement.
- Different tones: ask for the same post in two or three tones and pick the closest, then edit.
- Shorter: ask for a version that fits the tightest network you are posting to.
- Hashtags: ask for a short list of relevant hashtags, then remove any that are too broad to be useful or unrelated to the brand.
- Grammar and clarity: useful for a final pass, especially when the writer is working in a second language.
- Risk check: ask what could go wrong with the post: claims that need evidence, missing disclosures, insensitive timing, accessibility gaps.
What still needs a human
AI saves time on first drafts and variations. It does not replace the person who knows the client. Before any AI-assisted post is scheduled, a human should check:
- Facts: every name, date, price and link matches the brief and the real world.
- Claims: nothing promises a result, compares with a competitor or makes a health, financial or environmental claim the client cannot support.
- Context: the post still makes sense next to the news of the day and the rest of the calendar.
- Voice: it sounds like the client, not like an assistant.
- Accessibility: images have alt text, hashtags with several words use capitals, and the post does not rely on emoji to carry meaning.
For posts that need client sign-off, the AI draft goes through approval like any other post. Our approval checklist is a good companion for reviewers.
A brief template you can copy
Paste this into your notes and fill it in for each post:
- News: what is happening, with names, dates, times and prices
- Audience for this post:
- Why they should care:
- Action:
- Must include: links, disclosures, hashtags
- Avoid: words, claims, topics
- Networks:
Doing this in PostRoost
The AI writing assistant sits inside the composer. You describe the post as a brief, and it writes a general version plus one per selected network, within each network’s character limit and link rules, using the client’s brand voice from the client’s settings (who they are, who they talk to, words to use and avoid, hashtags they always use). From there you can ask for hashtags, three tones, a more engaging version, a grammar fix, a shorter cut or a risk check, which lists possible issues without changing your text.
The assistant uses Anthropic’s Claude models and is instructed not to invent facts, prices or dates that are not in your brief. It can still be wrong, so suggestions are only inserted when you choose them, and nothing is published without a person scheduling it or approving it. The Free plan includes five assistant requests a week per agency; paid plans are unlimited. See pricing for the details.
The short version
Give the assistant what you would give a freelance copywriter: the facts, the audience, the reason, the action and the rules. Set the brand voice once per client. Ask for per-network versions and specific improvements. Then read every word before it goes out.
