Every social media team adopts AI captions the same way. You open a tool, type a rough idea, hit generate, skim four options, pick the least robotic one, and post. It saves maybe ten minutes per post. It also produces captions that could belong to any brand in any country selling any product. After about three weeks your followers can feel the difference between the things you used to write and the things the machine writes, even if they cannot name what changed.

The problem is not that AI writes badly. Modern language models write fluent, correctly punctuated, hashtag-friendly prose on the first attempt, most of the time. The problem is that fluency is not voice. Voice is the set of specific, slightly inconvenient choices that make a sentence recognisably yours: the words you always use, the words you refuse to use, how long your sentences run, whether you open with a question or a claim, whether you ever break a line for effect. A model with no input about those choices will not guess them. It will average them. And the average of every brand voice on the internet is a flat, pleasant nothing.

This post is the workflow we use to fix that. Five steps, about ninety minutes to set up, then roughly four minutes per post after that: a brand voice brief, a caption bank mined from your own winners, a prompt structure that produces usable drafts, platform-specific rules, and a short human edit pass. There is also a section on measuring whether any of this is working, because a caption that sounds great and does nothing for your numbers is still a failed caption.

The goal is not to make AI sound human. The goal is to make AI sound like a specific human: the person your audience already trusts.

Why AI captions sound like nobody

Start with what is actually happening inside the model. When you ask for an Instagram caption about a new product launch with no other instructions, the model is doing pattern completion over an enormous pile of marketing text. The statistically safest completion of a marketing caption is the one that looks most like the middle of every marketing caption ever written. That is why the output arrives with a question mark in the first line, three emoji in the middle, a benefit list, a call to action, and a question addressed to nobody at the end.

There are four practical consequences you can observe in your own drafts.

• The model has no facts about you. It does not know your refund policy, your founder's name, the city your warehouse is in, or that your best-selling SKU is actually the unpopular colour. So it substitutes category-level claims: premium quality, effortless, designed for you. Those phrases are true of nothing.
• The model has no memory of your last forty posts. It cannot see that you already used "game changer" twice this month, that you never use exclamation marks, or that your audience is in Karachi and posts in a mix of English and Urdu transliteration.
• The model optimises for sounding correct. Correctness and personality pull in opposite directions. Personality lives in the deviations: a short punchy sentence after three long ones, a lowercase opening, a deliberately informal word in a formal category.
• The model has no stake in the outcome. It will not lose a customer if the tone is wrong for a service outage, and it will not gain one if the tone is right for a launch. Only you know which posts are which.

None of this means the tool is useless. It means the tool needs input. Everything in the rest of this post is about feeding a model enough of your real material that its output stops regressing to the mean.

The seven tells readers already recognise

You have seen machine captions in the wild. In blind tests with our own team, the tells came up again and again. Learn them so you can hunt for them during the edit pass.

1. 1The opening question nobody would actually ask out loud, such as "Ready to transform your morning routine?"
2. 2The closing question nobody answers, such as "What do you think? Let us know in the comments!"
3. 3Stacked abstract nouns: clarity, synergy, innovation, empowerment, seamless experiences.
4. 4Power verbs doing nothing: unlock, elevate, supercharge, revolutionise, unleash.
5. 5Three or more emoji placed as decoration between sentences rather than as punctuation.
6. 6A hashtag block that reads like a keyword dump, thirteen tags deep, half of them irrelevant.
7. 7A benefit claim with no number, no name, and no evidence attached.

If your draft contains three or more of the seven, it is not ready. Do not post it and hope. Run it through the edit pass later in this post instead.

The five failure modes of AI caption output

Before building the workflow, name the ways drafts actually fail. Each failure has a different fix, so a single generic instruction like "make it sound more like me" rarely helps. Diagnose first, then edit.

1. 1Factual vacuum. The caption is fluent and says nothing verifiable. Fix: feed the model real raw material — a spec, a quote, a number, a customer sentence — and require at least one specific detail in every draft.
2. 2Voice drift. The sentences are fine but the personality is gone. Fix: the brand voice brief in the next section, plus two or three of your own published captions as few-shot examples.
3. 3Structure collapse. Every caption follows hook, three benefits, call to action in that exact order. Fix: vary the structural instruction per post, rotating through five or six templates instead of one.
4. 4Platform mismatch. One caption gets pasted to seven networks with a hashtag swap. Fix: platform rules as a separate step after the draft, not as an afterthought.
5. 5Tone miscalibration. Cheerful language on an announcement about a price change, or corporate language on a joke. Fix: a post-type label — launch, education, community, apology, promotion — assigned before generation.

Here is a quick self-check you can run on any draft in under a minute.

• Does it contain at least one fact only we could write?
• Would a competitor be able to post it unchanged?
• Does it match how this particular post type should feel?
• Is the first line doing work, or is it throat-clearing?
• How many of the seven tells appear?

A caption you cannot defend in a client meeting is not on-brand. It is merely not off-brand, which is a much lower bar.

Step 1: Write a one-page brand voice brief

This is the ninety-minute step that everything else depends on. One page. Not a forty-slide brand guideline deck. A document short enough that a model can take it as context on every generation and a new hire can read it in six minutes.

Work through five blocks in order.

Block one: who you are talking to

Write three sentences about the actual reader, not the demographic. Include where they read (phone, in bed, on a commute), what they are sceptical about, and what they want to be able to say after seeing your post. For a B2B SaaS account the reader might be: a marketing manager at a 10-person company in Lahore, checking LinkedIn at 9am with a coffee, sceptical of anything that promises results without a number, who wants to forward the post to their founder with a one-line justification.

Block two: personality in dimensions, not adjectives

Adjectives like "friendly and professional" are useless because every brand claims them. Use three sliders with named poles instead, and write where you sit and where you deliberately do not sit.

• Confident versus humble: we are confident about methods, humble about guarantees.
• Playful versus serious: playful in community posts, serious in anything involving money or outages.
• Dense versus spacious: dense on LinkedIn, spacious on Instagram captions.

Then add two words you actively avoid. Ours were "game changer" and "dive in". The negative space defines a voice faster than the positive space does.

Block three: vocabulary rules

List the words and phrases you always use and the ones you ban. Include how you refer to your own product (never "our platform" in captions, always "PostHub"), how you refer to the reader (you, never "users"), and whether you ever use the word "content" out loud.

Block four: mechanics

Capture the boring but decisive details: sentence length preference, whether you use the serial comma, emoji policy, hashtag count and placement, line break habits, capitalisation of headings, whether you ever start a caption with a lowercase letter.

Block five: proof habits

Describe how you make claims. Do you always attach a number? Do you quote customers by name? Do you admit limits? A voice is partly a policy about evidence.

Here is a template you can copy and fill in. Keep the whole thing under 400 words so it stays affordable as prompt context.

BRAND VOICE BRIEF — [brand] — v1.2

READER: [three sentences, including where and when they read]
SLIDERS: confident/humble = 70/30 about methods, 0/100 about guarantees
         playful/serious = 60/40 community, 10/90 money + outages
         dense/spacious = dense on LinkedIn, spacious on IG
SAY: [8-12 words and phrases you use weekly]
NEVER SAY: [6-10 banned words, including category cliches]
REFER TO: product = [name], reader = you, competitors = never named
SENTENCES: mostly 8-18 words, one short punch per caption allowed
MECHANICS: [emoji policy] | [hashtag count] | [line break habit] | [capitalisation]
PROOF: every claim carries a number, a name, or an admission of a limit
POST TYPES: launch | education | community | behind-the-scenes | promotion |
            apology — each has its own tone note of one line

Save this somewhere you can paste it in seconds. In PostHub you can keep it attached to your workspace so every caption generated in the composer starts from your brief rather than from scratch.

Voice brief checklist

• Under 400 words, readable in six minutes
• Contains at least eight banned words, not just preferred words
• Names three post types with their own one-line tone notes
• Includes mechanics: emoji, hashtags, line breaks, capitalisation
• Written in the voice it describes, not corporate boilerplate
• Reviewed after 60 days using comments as evidence of how you actually land

Step 2: Build a caption bank from your own winners

Generic output is a modelling problem, and the cheapest fix for a modelling problem is better examples. Your own published posts are better examples than anything a model will invent, because they have already been tested by your audience.

Mine ninety days of your own history

Pull the last ninety days of posts and sort by the metric that matches the post's job. Reach for awareness posts, saves and shares for educational posts, comments for community posts, clicks for promotional posts. If your analytics platform makes this easy, use it — inside PostHub the top-performing posts view already sorts by engagement and lets you filter by account and date range, so the export takes about two minutes.

Take the top 30 to 50 captions. Do not judge them on taste. Judge them on numbers.

Extract patterns, not phrases

You are looking for structure, and there are five things worth extracting.

1. 1Opening lines. Copy the first sentence of every winner into one list. You will find two or three shapes recur: a claim, a number, a confession, a direct instruction. Those are your hooks.
2. 2Length distribution. Measure the winners. Most teams discover their real working length is far narrower than they assumed — often 90 to 140 characters for Instagram, or 600 to 900 characters for LinkedIn. Write the range down.
3. 3Proof devices. Count how often winners include a number, a customer quote, a before-and-after, a named person, a specific place. That ratio is your evidence budget.
4. 4CTA types. Separate "comment this word", "save for later", "link in bio", "no CTA at all". Note which post types use which. A lot of teams call to action on every single post and then wonder why comments dried up.
5. 5Format habits. Line breaks, list style, whether the caption repeats the on-screen text, whether hashtags appear at all.

Store the bank as a numbered list. When you generate, you will hand the model two or three of these as examples rather than describing what you want in the abstract.

Describing a voice gets you a paragraph about a voice. Showing a voice gets you a caption in that voice.

Keep the bank honest

Prune it every quarter. A caption that performed in January because of a trending audio is not a permanent template. Keep roughly forty entries, refresh ten each quarter, and date-stamp each one so you know how old the pattern is.

Caption bank checklist

• 30 to 50 winners from the last 90 days, sorted by the right metric
• Opening lines collected in one list, with the three recurring shapes named
• Measured length ranges for Instagram and LinkedIn separately
• Proof-device ratio counted, then written into the voice brief
• CTA types mapped to post types
• Dated and pruned quarterly

Step 3: The prompt architecture that produces usable drafts

A good caption prompt has eight parts. Most people write three of them and then blame the model for the missing five.

1. 1Context line. Who you are and who you are writing for, one sentence.
2. 2Voice brief. The document from step one, pasted in or referenced.
3. 3Raw material. The actual facts: the number, the quote, the product detail, the announcement. This is the part people skip, and it is why drafts come back empty of specifics.
4. 4Post type. Launch, education, community, promotion, apology — one word that changes the tone more than any adjective.
5. 5Platform and length. Name the network and the character budget.
6. 6Structure instruction. The shape you want: open with a number, end without a CTA, use a three-line list.
7. 7Few-shot examples. Two of your own winners from the caption bank.
8. 8Negative constraints. The banned words and the seven tells, explicitly.

Here is a working prompt. Replace the bracketed parts and it will produce three usable drafts instead of four forgettable ones.

Write 3 Instagram captions for [brand].

READER: [one sentence about who reads and when]
VOICE: [paste brief, or: plain sentences, confident about methods, never
       guarantees, no exclamation marks, product is always called [name]]
POST TYPE: education
FACTS (use at least one, verbatim where possible):
- [number, e.g. "teams using a shared approval queue cut revision rounds
  from 4.2 to 1.6"]
- [source, e.g. "survey of 340 social media managers, August 2026"]
- [constraint, e.g. "only applies to accounts on the team plan"]
TARGET: 110-140 characters, first line must work as a hook, no CTA,
        at most one emoji, max 3 hashtags at the end.
STRUCTURE: open with the number, one sentence of context, one line that
           tells the reader what to do with it.
EXAMPLES OF OUR VOICE (match the rhythm, not the content):
1. [paste winner]
2. [paste winner]
NEVER: game changer, dive in, unlock, elevate, seamless, "what do you think",
       stacked emoji, hashtag blocks over 3 tags, rhetorical questions.
Give me 3 options with different openings.

Why examples beat adjectives

If you write "write in a warm, witty, confident tone", the model will produce something that matches its internal average of those three words. If you paste two captions you actually published, the model copies measurable properties instead: average sentence length, where the line breaks fall, how much hedging appears, whether the punchline comes before or after the proof. One example is worth roughly forty adjectives.

Run a three-draft loop, not a one-draft hope

Ask for three options with deliberately different openings. Pick the best. Then, before you post, apply the edit pass and save one improved line back into your caption bank. Over a month you are not just posting — you are retraining your own process. The bank gets better, which makes the next prompt better, which makes the drafts better.

Two settings matter. Keep temperature low for factual posts so claims stay tight, and raise it for community or humour posts where you want real variation. Set a length control before generating rather than trimming afterwards — cutting to a character budget is how you lose the specific detail that made the caption work.

Prompt architecture checklist

• All eight parts present, especially raw material and negative constraints
• At least two real examples from your caption bank
• Character budget and platform named explicitly
• Post type assigned before generation
• Three drafts requested with different openings
• One improved line recycled into the bank after every post

Step 4: Platform rules — one idea, seven dialects

The single most expensive shortcut in social media is writing one caption and pasting it everywhere. Each network has a different visible-text budget, a different scanning behaviour, and a different reason for showing your post to people. Adapt the idea, not just the hashtags.

Instagram

The first line is the only line most people see before the truncation point, so it has to carry the hook alone — think of it as a headline with a caption attached. Keep the total in your measured range, usually 90 to 140 characters for feed posts, and use line breaks to create a readable shape rather than a wall. Save-bait works: tell people why they would come back to it. Hashtags are a discovery aid, not a ranking lever; three to five well-chosen tags beat twenty generic ones, and they belong at the end or in the first comment.

LinkedIn

Open with a line that survives the "see more" fold, usually under about 210 characters including the reader's name and network context. Then give the post air: short paragraphs of one to two sentences, because dense blocks are scanned and skipped. Hashtags are largely optional now; one or two topical tags are plenty and many high-performing posts carry none. Numbers and first-person experience outperform polished brand language. Do not paste a link in the body if you can put it in the first comment instead, and never open with "I am excited to announce".

TikTok

The caption is not the star — the video is — so the caption's job is keyword clarity and a nudge to watch again or check your profile. TikTok's search treats caption text as a ranking signal, so put the plain-language topic in the caption rather than hiding it behind cleverness. Keep it short, skip the hashtag soup, and assume people read it only after they have already watched.

YouTube Shorts

Treat the description like a search field. Lead with a sentence containing the exact phrase a person would type, then add context. The first 150 characters carry the most weight in browse and search results, so do not open with a joke at the expense of clarity. A pinned comment can carry the links that the description buries.

Threads

Threads rewards conversation, not broadcast. Write the way you would reply to a smart friend: a specific observation, a small claim you are willing to defend, an invitation to disagree. Hashtags are not a growth mechanism here. Posts that read like a Twitter-length thought perform better than posts that read like a press release, and replying to your own post with the follow-up detail often keeps a thread alive longer than one long post.

Facebook

The audience skews toward personal context, so warmth and directness beat polish. Links are acceptable and often expected, but put a sentence of your own above the link rather than dropping it bare. Longer captions do fine when they tell a story, especially for local and community audiences. Groups behave differently from pages: in a group, answer questions and skip the marketing framing entirely.

Pinterest and Bluesky

Pinterest descriptions are keyword surfaces — write a plain sentence describing what the pin shows and who it is for, then add a short benefit line. Bluesky behaves like a text-first network: concise, opinionated, no hashtag dependence, and a reply under your own post is better than a follow-up post.

Here is the adaptation checklist we run after the core draft exists.

• First line rewritten to survive that platform's truncation point
• Character count inside the platform's measured range
• Line breaks and paragraph shape matched to how people scan there
• Hashtag count set to that platform's habit, not copied from elsewhere
• Link placement chosen for that platform rather than identical everywhere
• Post type tone checked: does a promotion still belong in this feed today?
• On-screen text and caption checked for repetition in video formats

Cross-posting the same sentence seven times is not efficiency. It is seven chances to look like you are not paying attention.

Step 5: The 60-second human edit pass

This is the step teams skip when they are rushed, and it is the step that separates captions people trust from captions people scroll past. Sixty seconds, seven questions, in order.

1. 1Is there one fact only we could write? If not, add a number, a name, a place, or a limit. Specificity is the cheapest authenticity available.
2. 2Does the first line earn the second? Cut any opening that restates the obvious. "We're thrilled to share some exciting news" is a sentence that could be deleted with no loss.
3. 3Would I send this to a client as a message? Read it aloud, in your head, as if texting a person who pays you. If it sounds like a broadcast, rewrite it as speech.
4. 4Is the claim defensible? Every superlative needs evidence or a qualifier. "Fastest" needs a benchmark. "Loved by thousands" needs a number you can source.
5. 5One call to action, maximum. Pick the action you actually want. Two competing asks cancel each other.
6. 6Count the tells. Run the seven tells from earlier. Three or more means another pass.
7. 7Check the mechanics against the brief. Emoji, hashtags, line breaks, capitalisation — the boring stuff readers notice when it is wrong.

The concrete rewrite example

Here is a real transformation from a draft we received, stripped of brand names.

Before:

🚀 Ready to transform your social media game? Say goodbye to scattered
tools and hello to effortless scheduling! 🎉✨

Our all-in-one platform helps you plan, publish and analyse — all in one
place. What do you think? Let us know below! 👇

#socialmedia #scheduling #marketing #growth #smallbusiness #tools

After the edit pass:

We tracked 340 posting schedules last quarter. The teams that published
the same number of posts but cut their tool-switching time in half
finished their week two hours earlier.

One queue. One approval. One analytics view. That is the whole product.

Start with your busiest account: https://posthub.shorix.io

Same topic, same platform, roughly the same length. The difference is that the second version contains a number, removes four rhetorical questions, drops eleven hashtags, and ends with a single action instead of an invitation to feel something. It could not be posted by a competitor, because the number belongs to you.

Where the sixty seconds actually go

Teams say they cannot afford a minute per post. In practice the minute is already spent — as a panic edit after publishing, or as a reply explaining what the post meant. Front-load it.

Edit pass checklist

• One specific fact survives from draft to publish
• First line deleted if it only announced that something is coming
• Read aloud test passed without cringe
• Every superlative has evidence or a qualifier
• Exactly one call to action
• Three or fewer of the seven tells
• Mechanics matched to the voice brief

Mistakes and fixes from teams that have done this already

• Mistake: asking the model to "sound more authentic". Fix: give it material instead. Authenticity is a by-product of specificity. Paste the customer quote, the shipment delay, the number of signups.
• Mistake: keeping one master prompt forever. Fix: version it. Date your prompts the same way you date your voice brief, and retire one every month that no longer reflects how you write.
• Mistake: generating captions before you know the post's job. Fix: label every post — launch, education, community, promotion, apology — before you open the composer. The label changes the tone more than any editing pass will.
• Mistake: letting five people edit in five directions. Fix: one owner per caption, with an approval step that collects comments rather than messages. This is the difference between an edit and an argument.
• Mistake: measuring only likes. Fix: track saves, shares, profile visits and follower change per post, which are the signals of whether the caption persuaded anyone. The next section covers this.
• Mistake: using AI for everything, including the hard posts. Fix: keep human-only categories. Apologies, crisis updates, pricing changes affecting existing customers, and anything about a sensitive cultural moment should be written by a person, reviewed by a second person, and published on purpose.
• Mistake: never reviewing what the model produced. Fix: a monthly audit of twenty published captions, scored against the seven tells. If your tell-score is rising, your voice brief is stale.

How to tell whether your AI captions are actually working

A workflow that produces captions you like but that do not perform is a comfort routine, not a strategy. Measure it the way you would measure any other change.

Choose metrics by post job

Do not use one metric for all posts. Assign a primary metric per job and read only that number for the first two weeks.

• Awareness posts: reach and non-follower share of reach
• Educational posts: saves and shares, because these are intent signals
• Community posts: comment count and comment sentiment, not just volume
• Promotional posts: link clicks and click-through rate
• All posts: profile visits per 1,000 impressions, which shows whether the caption made people curious about you

Run a simple split for thirty days

For thirty days, alternate your approach by weekday. Posts on Monday, Wednesday and Friday get the full workflow: voice brief, caption bank, structured prompt, edit pass. Posts on Tuesday and Thursday get the old habit: generate, skim, post. Keep the topic mix roughly equal so the comparison means something.

Then read three numbers side by side: median saves per post, median profile visits, and follower change across the period. You are looking for direction, not statistical significance — this is a small sample and you should treat it as evidence rather than proof.

Watch for the substitution effect

The most common bad outcome is not worse performance — it is easier posting. AI captions make it simple to publish more, and publishing more dilutes quality until the median drops even if your best posts still land. If volume rises 40 percent and median engagement falls, you have found your ceiling. Cut back to the volume where quality holds.

Use the loop, not the dashboard

Analytics should change the next prompt, not just fill a report. Once a month, take your three best and three worst captions of the month, put them side by side, and write one sentence about the difference. Feed that sentence into the voice brief or the negative constraints. In PostHub you can open the analytics view, sort the period by engagement, and jump straight from a post to its caption so the comparison takes a couple of minutes rather than an afternoon. If you want to see how the reporting is structured, the PostHub overview walks through the composer and analytics in the same workspace.

The point of measuring captions is not to produce a chart. It is to make next month's prompt smarter than this month's.

Putting the whole workflow together

Once the pieces exist, the daily routine is short. Here is the full sequence from idea to published post, with realistic timings for a single-person team.

1. 1Decide the job — post type and primary metric, 30 seconds.
2. 2Collect raw material — the number, quote or detail you will use, 2 minutes. This is the step that cannot be automated honestly.
3. 3Generate three drafts with your versioned prompt, 30 seconds.
4. 4Pick and edit using the sixty-second pass, 60 seconds.
5. 5Adapt per platform with the checklist, 60 to 90 seconds per network.
6. 6Route for approval if you are not the final approver — one queue, not a chat thread.
7. 7Schedule and let the publishing window handle the rest.
8. 8Review weekly — pull the top three posts, note one pattern, update the bank.

Total for a post going to three networks: about six minutes of writing and adaptation, plus whatever your review step requires. Compare that with the fifteen to twenty-five minutes most teams spend when every caption starts from a blank box.

Teams that run this inside PostHub keep the voice brief, the caption bank and the analytics in one place, so the loop between "what performed" and "what we generate next" does not require exporting spreadsheets. If you are evaluating the tool for a team, the pricing page shows which plan includes multi-user approvals and shared asset libraries.

When the workflow should not be used at all

Some posts should be written by a person from a blank page with no model in the loop. Keep a standing list and share it with everyone who publishes.

• Anything that admits a mistake, an outage, a delay or a refund issue
• Anything touching a sensitive cultural, religious or political moment in any market you operate in
• Any claim about health, money, legal outcomes or guaranteed results
• Personal notes from a founder about their own experience
• Anything where the joke depends on reading the room today

If a post matters enough that the wording could cause a problem, the wording deserves a human and a second human.

Key takeaways

• AI captions fail because of missing input, not missing talent — feed facts, examples and constraints, not adjectives.
• Write a one-page voice brief under 400 words with banned words, mechanics and post-type tone notes; revisit it every 60 days.
• Mine 30 to 50 of your own top performers and keep them in a caption bank you draw two or three examples from on every generation.
• Use an eight-part prompt: context, voice, raw material, post type, platform and length, structure, few-shot examples, negative constraints.
• Ask for three drafts with different openings, then run a sixty-second edit pass before anything goes out.
• Adapt per platform after the core draft: first line, length, line breaks, hashtags and link placement are all different everywhere.
• Measure by post job — saves for education, profile visits for curiosity, clicks for promotion — and feed one insight back into the prompt every month.
• Keep a hard list of post types that never get generated: apologies, crises, sensitive moments, personal notes.

FAQ

Does using AI make captions sound the same across every brand?

Only if you give every brand the same input. Model defaults are shared, but the brief, the raw material and the examples you provide are not. Two teams working from different voice briefs and different caption banks will produce visibly different output even from the same model. The homogeneity people complain about is usually a symptom of teams all skipping the same step.

How long should the voice brief be?

Under 400 words. Long enough to include banned words, mechanics and post-type notes, short enough that it can sit inside every prompt without crowding out the actual facts of the post. If your brief needs twenty pages, it is a brand book, and a brand book does not fit in a prompt — extract the two pages that change how sentences are written.

Should I tell my audience when captions are AI-assisted?

Be honest about what the audience actually cares about: accuracy and usefulness. Nobody needs a disclaimer on a scheduling tip. Do disclose when a caption makes a factual claim you did not verify, and never let a model invent quotes, statistics or customer stories. If your process puts a human edit between the draft and the publish button, you are already responsible for every word.

What is a realistic time saving?

For a team already publishing five to ten times a week, expect roughly 40 to 60 percent less time per caption once the brief and bank exist. The setup costs about ninety minutes up front and the prompt template costs another thirty. Most teams recover that within the first week. The saving shrinks if you skip raw material collection, because then you spend the time backediting vague claims.

Can this work for a team where several people write captions?

Yes, and it works better. The brief and the bank become shared references rather than one writer's instinct, so quality stops depending on who is at their desk. Pair the shared prompt with a single approval queue so edits are collected in one place instead of scattered across chat. That combination is what makes tone consistent across five writers.

How do I stop the output from drifting back to generic after a few weeks?

Drift almost always comes from two causes: a stale caption bank and a prompt nobody has versioned. Refresh ten examples in the bank each quarter, date every prompt, and run the monthly twenty-caption audit against the seven tells. If the tell-score climbs for two months in a row, the brief no longer describes how you write, and the brief is what needs rewriting — not the individual captions.