AI for social media

Using AI for social media.

An assistant with no context writes the average of the internet. This is how to give it yours, how to prompt it properly, and how to make it get better as your account grows.

Updated August 2026 / 16 min read

The short answer

To use AI well for social media, stop asking it for posts and start giving it context. Write one file that says who you are, who you are talking to, what you have already published, and what your numbers actually look like. Paste that file into every session. Ask for options rather than answers, and judge them yourself. Then feed the results back after you post, so the next draft is written against your evidence instead of the average of the internet.

The honest scope

What AI is actually for.

More than eight in ten creators now use these tools somewhere in their process, so the tool itself stopped being an advantage roughly two years ago. What separates accounts now is not whether a model was involved. It is whether the person using it knew what to hand over.

The split is cleaner than most guides admit. A model is extraordinary at producing range, structure and second opinions on material you already have. It is useless at deciding what is worth saying, and actively dangerous when it is asked to invent a life you have not lived.

Range, on demand

Twelve different angles on the same idea in thirty seconds. You were never going to write twelve. You only need to be able to recognize the good one, which is a much cheaper skill than generating it.

Structure from a mess

You talk about your idea for four minutes into your phone. It comes back as a script with an opening, a middle that earns the ending, and a note on the sentence you should lead with.

Memory across your own back catalogue

You cannot hold two years of your own posts in your head. Paste them in and ask what your best twenty have in common. The answer is usually not the thing you assumed.

A second pass that is not flattering

The most valuable question you can ask is not write this for me. It is read this as someone who does not follow me and tell me where they leave.

Translation between formats

One video becomes a carousel, a caption, a newsletter opener, and three follow up ideas. This is genuine compounding and it costs you almost nothing.

The boring middle

Subtitles, alt text, descriptions, chapter titles, the fourth variation of a caption for a different platform. Work that has to happen and that nobody has ever been rewarded for doing personally.

And the four things it quietly costs you

Deciding what you think

A model will happily produce a point of view. It will be the median point of view, held by nobody, and your audience follows you for the opposite of that.

Taste

Knowing which of the twelve hooks is actually good is the entire job. That judgment is the thing you are being paid for, and it is the one part that does not transfer.

Experience you have not had

The fastest way to lose an audience is a story that did not happen to you. It is also the easiest thing in the world to generate, which is exactly why it is a trap.

Volume for its own sake

Posting five times more of something nobody wanted is not leverage. It is the same problem, five times, in front of an algorithm that now has five data points saying you are skippable.

The failure mode

Why generic output fails.

A language model predicts the most likely continuation of the text in front of it. Most likely means most common. Most common means the thing a thousand accounts in your niche already posted this month. That is not a flaw in the technology, it is the technology working exactly as designed, and with no information about you the most likely thing is the only thing it can reach for.

Given nothing, it returns the statistical middle of everything it has ever read. The statistical middle of social media advice is what your audience has been scrolling past for two years.

Nothing given

Write me a hook for a Reel about morning routines.

You have been doing your mornings wrong your whole life. Here are 5 habits that changed everything for me.

Correct, competent, and interchangeable with every other video on the subject.

Context given

Here is my context file. I am a night shift nurse who posts about being awake when everyone else is asleep. My audience is other shift workers who are sick of 5am routine content. Give me 12 openings for a video about what my morning actually looks like at 7pm, and include at least three that mock the genre.

My morning starts at 7pm and no productivity guru has ever made a video for me.

Same model, same minute. The difference is entirely in what it was told.

The two bills that arrive

The first is the audience. In Sprout Social's Q1 2026 survey of 2,250 people, half of Gen Z respondents said they had already blocked, muted or unfollowed a brand or creator over content that felt machine made, and 56 percent said they see that kind of content often. Two thirds said they are more selective about what they engage with than they were a year ago.

The second is distribution, and it follows from the first. Nobody is scanning your captions for a machine. Your audience is, and the ranking systems are watching your audience. A post people scroll past is a post that stops being shown, and the cause of the scroll never appears in your analytics.

Instagram is explicit about the neighboring case. Since 30 April 2026, accounts that mostly repost work they did not create, without meaningfully transforming it, are no longer eligible for recommendation to people who do not already follow them. The principle underneath it is the one worth carrying: the platform is trying to find things only you could have made.

The core idea

Context is the whole job.

The people building serious products on top of these models stopped calling this prompt engineering. They call it context engineering, because once real systems were shipped it turned out that most of the quality lived in what the model was given rather than in how the request was phrased. A sharp prompt on empty context still fails. An ordinary prompt on rich context usually works.

For a creator, context is five layers. You can write all of them in ten minutes, and nearly every disappointing output you have ever had is one of them missing.

Identity

Who you are, what you actually do for money, and what this account is supposed to get you. There is a real difference between an account trying to reach ten thousand followers and an account trying to book four clients a month, and almost every piece of advice a model will give you assumes the first one.

TELL If the draft could have been posted by any of a thousand accounts in your niche, layer one is missing.

Audience

Not demographics. One person, their problem stated in their own words, what they have already tried and what they are tired of hearing. Age brackets and interest categories are what an ad platform needs. A writer needs to know what makes this person feel understood.

TELL If the draft explains something your audience already knows, layer two is missing.

Your body of work

Ten to fifteen real posts, pasted in full. Your best, and at least two that failed. This is the single highest leverage thing on this page and almost nobody does it. Telling a model you are witty and down to earth gives it nothing, because every brand on earth describes itself that way. Showing it three captions you actually wrote gives it your sentence length, your punctuation habits, the jokes you make, the way you open, and the fact that you never use exclamation marks.

TELL If the draft sounds like a brand rather than a person, layer three is missing.

Your numbers

Average views over your last twenty posts. Your best and worst, with what you think happened. Saves and sends if you can get them. Without this, a model has no idea what good means for you, and it will cheerfully call a result a win or a failure with no basis for either.

TELL If the advice would be identical for an account with a hundred followers and one with a hundred thousand, layer four is missing.

Your constraints

Where you can film, how much time you really have on a Tuesday, what gear exists, what you will not talk about, and any brand rules you are bound by. Ideas you cannot execute are not ideas, they are a way of feeling productive while nothing gets posted.

TELL If the plan needs a second person, a studio, or six hours you do not have, layer five is missing.

The artifact

Your context file.

One document. Not a prompt, not a chat. A file you own, that you paste at the start of a session or keep pinned in a project, and that you update on the first of the month. Everything else on this page is downstream of it.

Creator context file

WHO I AM
Handle:
What I actually do:
What this account is supposed to get me (be specific: clients, sponsors, a launch, a number):
What I will never post:

WHO I TALK TO
The one person:
Their problem, in their words:
What they have already tried:
What they roll their eyes at:

MY VOICE
Three posts I would show a stranger: [paste them in full, not links]
Two posts that flopped: [paste them]
Words I use a lot:
Words I never use:
How I open a video:
How long my sentences run:

MY NUMBERS (update this monthly)
Followers:
Average views, last 20 posts:
Best post and why I think it worked:
Worst post and why I think it failed:
Typical saves / sends:

MY CONSTRAINTS
Where I film:
Time per week:
Formats I can actually make:
Formats I cannot:
Anything I am contractually not allowed to say:

Fill in what you know. A half finished version beats what almost every account is working with, which is nothing at all.

The craft

Nine rules for prompting.

With the file in place, prompting stops being a trick and becomes a way of asking. These nine hold across every model worth using, and most of them are about giving up the fantasy that the machine will do the judging.

Ask for the job, not the vibe

Adjectives are the weakest instruction you can give. A job has a shape you can check the output against.

WeakWrite a fun, engaging caption.

BetterWrite a caption under 40 words that ends on the question my audience argues about in the comments.

Show, do not describe

Three of your own posts pasted in full will do more for the voice than a paragraph of self description ever will. This is the difference between telling a session musician you want it soulful and playing them the record.

WeakMatch my tone: casual, witty, a bit sarcastic.

BetterHere are four captions I wrote. Match the rhythm and the sentence length, not the topics.

Name the viewer, not the platform

Writing for Instagram produces content for nobody. Writing for a 26 year old who has tried three budgeting apps and quit all of them produces a line.

Ask for many and judge yourself

One answer invites you to accept or reject. Twelve answers put you in the editor chair, which is where you are actually good. Ask for range on purpose: different angles, not different wordings of the same angle.

Make it think before it writes

Ask for the plan first, approve the plan, then ask for the draft. A model that has already committed to a structure writes a better second half, and you catch a bad idea before it is dressed up in good sentences.

Front load the constraints

Length, format, what you can film, what you cannot say, deadline. Constraints given afterwards produce a rewrite. Constraints given up front produce a usable first draft.

Ban your own tells

Every model has house habits, and audiences have learned them. Keep a running list and paste it every time. It is the cheapest quality improvement available to you.

WeakMake it sound less like AI.

BetterNever use: delve, unlock, elevate, game changer, in today's fast paced world, the truth is, let that sink in. No sentence may start with the word Imagine. No rule of three.

Critique, do not regenerate

Pressing regenerate is a coin flip that throws away everything you just learned. Telling it that the second line is where you stopped believing it is a fix, and the fix carries into the next draft.

Keep the session, version the file

A long thread is working memory. Stay in it while you are on one job, and start fresh when you change jobs, not when you get a bad answer. The permanent memory is your context file, and it lives in a document you own rather than in a chat you will lose.

The library

Six prompts to steal.

Every one of these assumes your context file goes in where it says paste. They are long on purpose. A short prompt is a short brief, and you would not hand a short brief to an editor either.

The hook sprint

You have an idea and no opening line.

Prompt

You are helping me write the first three seconds of a short video.

Here is my context file:
[paste]

The idea, in one sentence:
[paste]

Give me 12 different openings. Vary the ANGLE, not the wording. Include at least one that opens on the result, one that names the mistake, one that starts mid story, one that argues with the common advice in my niche, and one that is a plain sentence with no technique in it at all.

After each, add a note of five words or fewer: who it stops, and why.

Then rank them for the audience described in my context file, and tell me which three you would keep if you could only keep three, and why the other nine lose.

Do not write the rest of the script yet.

WHY IT WORKS It asks for range instead of quality, forces a reason for each option, and stops before the part you should be doing yourself.

Voice note to script

You can explain it out loud but the page stays blank.

Prompt

Below is a transcript of me talking through an idea out loud. It is messy and it repeats itself.

[paste transcript]

Do not translate my words into marketing language. Keep my phrasing anywhere it is usable.

1. Tell me in one line what my actual point is.
2. Quote the single strongest sentence I said.
3. Rebuild it as a script of 130 words or fewer, in this order: strongest sentence first, then the setup, then the proof, then the turn. Label each part.
4. Flag every place where I claimed something I did not back up.

Keep contractions. Keep at least three phrases exactly as I said them. If the idea is not strong enough to carry a video, say so instead of fixing it.

WHY IT WORKS It makes the model an editor of your material rather than the author of new material. That is where nearly all of the quality is.

The pre-post critic

It is written and you want a real note before you film.

Prompt

Here is my script:
[paste]

Here is who it is for:
[paste the audience section of my context file]

Read it three times.

Pass 1: as someone scrolling who does not know me and does not want to watch anything. Mark the exact word where you would scroll, and say why.
Pass 2: as someone who already follows me. Mark where it gets boring for them specifically.
Pass 3: as someone who disagrees with the premise. Tell me the objection I never answer.

Then give me the one change that keeps the most people, and tell me what that change costs me in clarity, length, or honesty.

Do not compliment the script. Do not rewrite it.

WHY IT WORKS A model's default setting is encouraging. Naming the viewer, the exact moment and the cost of the fix is what turns that into an actual note.

The flop post mortem

Something underperformed and you want more than a guess.

Prompt

This post underperformed and I want to know why.

My last 20 posts averaged [X] views. This one got [Y], with [saves] saves and [sends] sends. Watch time fell off a cliff at [T] seconds.

Here is the script and the caption:
[paste]

Here is my context file:
[paste]

Give me three COMPETING explanations for the underperformance, ranked by how much of the gap each one would explain. For each: what evidence in my own analytics would confirm it, and what evidence would kill it.

Then tell me which explanation is testable in my next post, and exactly how to test it so the result means something.

Do not tell me to post more consistently, and do not tell me the algorithm is unpredictable.

WHY IT WORKS It forces competing explanations instead of one confident story, and makes the model tell you how to prove itself wrong.

One into five

Something worked and you want to compound it instead of repeating it.

Prompt

Here is a video that outperformed my baseline, with its numbers:
[paste script and numbers]

Here is the top comment on it:
[paste]

Give me five follow ups that are NOT the same video again:
1. the same point aimed at a different audience
2. the strongest counterargument, argued honestly
3. the version for someone one level further along than the original viewer
4. the story of why I believe this in the first place
5. the video that answers the top comment

For each: the opening line, the format, and what it is trying to earn (reach, saves, or trust). Match the voice in my context file. If one of the five would be weak, say so and give me a sixth instead of padding it.

WHY IT WORKS Repeating a winner burns it out. Extending it keeps the thing that worked and gives the audience a reason to stay.

The trend fit check

Everyone is doing the thing and you are not sure you should.

Prompt

Here is a trend I keep seeing:
[describe it, paste the audio name or a link]

Here is my context file:
[paste]

Answer in this order:
1. What is the trend actually rewarding: the audio, the format, the joke, the timing, or novelty that is already gone?
2. Which part of it transfers to my niche, and which part would look borrowed on my account?
3. Give me one version that keeps the transferable part and drops the rest.
4. One honest sentence: should I skip this?

You are allowed to tell me to skip it. If it does not fit the audience in my context file, say so plainly rather than finding a way to make it work.

WHY IT WORKS Most trend advice is do it faster. The useful question is which part of the trend is doing the work, and whether that part survives being moved into your niche.

The loop

Making it grow with you.

Here is the part that almost nobody does, and it is the part that compounds. By default an assistant has no memory of your account. Every session starts from zero, which means it is giving you the same advice in month six that it gave you in week one, while your account has moved on without it.

The fix is unglamorous. You keep a ledger, and you feed it back. One line per post, in the same shape every time, so that a month of posting becomes something a model can actually read.

One ledger line

2026-08-14 | reel | hook: names the mistake | 41s | talking head + b roll | 12,400 views (baseline 4,800) | 310 saves | 88 sends | note: posted 7pm, top comment asked about pricing

After every post

One line in the ledger

Not analytics. One line of plain text per post, in a document, in the same shape every time. It takes twenty seconds and it is the only thing that turns your account into training data. A month of these is worth more to a model than every general social media article ever written.

Every month

Paste the ledger back

Ask what your baseline is now and whether it moved. Ask which posts beat it by half again, and what those posts share that the others do not. Insist that a pattern is only named if at least three posts support it, otherwise you are being told a story about noise.

Every month

Retire something

The hard part of growing is not learning new rules, it is noticing when an old one stopped being true. Formats saturate. A hook that worked in March is a cliche by July because everyone copied it. Deleting a line from your context file is as important as adding one, and nothing will prompt you to do it.

Every quarter

Refresh the voice samples

Swap the example posts for your recent best. Your voice moves. If the model is imitating who you were a year ago, it will keep pulling you backwards into a version of yourself you have already outgrown.

Your baseline beats anyone else's benchmark

Four thousand views is a disaster for one account and a breakout for another. Every number you read in a general guide is an average of accounts that are not yours. A model that knows your last twenty posts can tell you which one you just had. A model that does not will confidently tell you the wrong one.

The monthly read

Here is my content ledger for the last 30 days, one line per post: date, format, hook type, length, views, saves, sends, note.

[paste]

1. What is my actual baseline this month, and did it move compared to the month before?
2. Which posts beat the baseline by more than 50 percent, and what do they share that the others do not? Only name a pattern if at least three posts support it.
3. What did I do more of this month that did not pay?
4. Name one thing to stop, and one thing to test twice next month.

If the data does not support a conclusion, say that instead of guessing. I would rather have one real finding than four confident ones.

The uncomfortable part

Most social media advice has never been tested.

There is a problem sitting underneath everything above. The model learned social media by reading the same coaches, threads and courses you did. It repeats their consensus with total confidence and no evidence, and consensus in this industry is mostly a measure of how often something has been repeated.

We went and checked. We pulled the concrete claims out of what working coaches actually teach, clustered them so we could see where they agree, then tested them against real performance data from a corpus of real posts. This is what came back.

The ClipScore rulebook, read 21 August 2026

175

Rules extracted

From what working coaches actually teach

Confirmed

Held up against real performance data

Contradicted

The data pointed the other way

157

Untested

Repeated everywhere, checked by nobody

Of the eleven claims the data argued with, five were about hooks, which is the single thing every coach is most certain about. And the largest number on this panel is the untested one. That is the honest state of the field: an enormous body of advice, repeated everywhere, checked almost nowhere.

We are not claiming the 157 are wrong. Untested means untested, and our own confirmed set is small enough that we say so on the panel rather than rounding it into a marketing number. The point is what it means for how you should use a model: treat every rule it gives you as a hypothesis with your name on it, not an instruction.

Testing one is not complicated. Run the same idea two ways, two weeks apart, change one variable, and write both results in the ledger. Six of those and you know more about your own account than any general guide can tell you, including this one.

The line

What to never hand over.

None of this is a moral position. These are the four places where handing the work over reliably costs more than it saves.

Your point of view

If you do not know what you think about your subject, no model can cover for that, and the gap is visible in about four seconds. Work out what you believe first. Then use the machine to say it better.

Your face and your voice

Synthetic presenters are cheap now and audiences are unusually good at spotting them. The moment someone suspects the person is not real, everything else you have said becomes suspect too.

Experience you did not have

A generated client story, a transformation that did not happen, a number you cannot show. This is the one that ends accounts rather than slowing them down, and it is the easiest possible thing to generate.

The last read

Nothing goes out unread. Not because a model will produce something offensive, but because it will produce something almost right, and almost right is what your audience learns to scroll past.

On disclosure, the research is blunt: the behavior audiences most want stopped is machine generated content posted without a label. If a model helped you sharpen something you were going to say anyway, most people treat that the way they treat spellcheck. If it generated the substance, say so. Being caught costs more than being honest ever did.

Today

Your first 30 minutes.

Write the context file

10 min

Use the template above. Do not polish it. Half filled in is already better than what most accounts are working with.

Paste in your work

5 min

Ten posts you are proud of and two that died, in full text, under the voice section. This is the step people skip and it is the step that does the work.

Find your baseline

5 min

Paste the view counts of your last twenty posts. Ask for your median, your outliers, and what the outliers have in common. Put the answer in the numbers section of the file.

Run the hook sprint

5 min

On the idea you already have in your notes. Keep three openings. Notice how different they are from what you would have got with a bare request.

Start the ledger

5 min

Backfill your last five posts in the one line format. Next month you paste it back and the loop starts paying you.

Next month, paste the ledger back and ask the four questions. That is the whole loop, and it is the difference between a tool you use and a system that knows you.

One note from us

Doing all of this automatically.

Everything above works. It is also a second job: a file to maintain, a ledger to keep, a monthly read to remember, and a body of advice to test one post at a time.

That job is the one we built ClipScore to do. It reads your actual account, so the context is your real posts and your real numbers rather than a document you keep up to date by hand. It remembers what worked for you specifically, holds it against the rulebook further up this page, and tells you when a rule stops being true for your account.

If you would rather run the loop yourself, the prompts above are yours and they cost nothing. If you would rather not, that is what the Studio is.

Open the Studio Or score a single reel first

Questions

The questions people actually ask.

Does using AI hurt your reach on Instagram?

Not for using it. Instagram's originality rules target reposting and aggregating other people's work rather than the tools you used to make your own. What does cost you reach is the audience response: Sprout Social's Q1 2026 survey found half of Gen Z have blocked, muted or unfollowed a brand or creator over content that felt machine made. Fewer people watching is a distribution problem, whatever caused it.

Will people be able to tell?

They can tell when the writing has no specifics in it. Generic structure, tidy rule of three, an opinion nobody could disagree with. They cannot tell when a real person used a model to sharpen something they were already going to say. The difference is not the tool, it is whether anything in the post could only have come from you.

What is the best AI tool for social media?

The model matters far less than the context you give it. Any current frontier assistant will outperform a better model that knows nothing about you. Choose one that lets you keep a persistent project or memory so your context file does not have to be pasted from scratch every session, and stay in it long enough that you learn its habits.

How do I make AI sound like me?

Paste your own writing, in full, and ask it to match rhythm and sentence length rather than topic. Give it two examples of what you sound like at your worst as well, so it knows the edges. Then keep a banned words list of the tells you notice, and paste that too. Description does almost nothing; examples do almost everything.

Should I tell my audience I used AI?

If it helped you write something you would have written anyway, most audiences treat that the way they treat spellcheck. If it generated the substance, the story or the face, say so. The thing consumers most want brands to stop doing is posting machine generated content without labeling it, and the trust cost of being caught is far higher than the cost of saying it upfront.

Can AI tell me what is trending right now?

Not on its own. A model's knowledge stops at its training cutoff, so anything it tells you about this week's feed is either from a search tool it is using live or an educated guess dressed as fact. Ask where the answer came from. If it cannot show you a source or a live view of the platform, treat it as a hypothesis.

How long before this actually starts working?

The prompts help immediately. The loop takes about a month, because that is how long it takes to have enough ledger lines for a pattern to be real rather than a coincidence. The accounts that get value from this are the ones still updating the file in month three.

Can AI run my whole content calendar?

It can draft one. It cannot decide one, because deciding requires knowing what you are willing to be known for, which is the part it has no access to. Use it to fill a structure you chose, and to notice when the structure stops working.

What we read

Written by the ClipScore team. The rulebook figures are our own, read out of our database on 21 August 2026, and they will move as the corpus grows. Everything else is linked above. Nothing here guarantees a result: the loop makes your decisions better informed, it does not make the platform owe you anything.