How-To Guide
What are the best Claude prompts for social media posts?
Eight copy-paste Claude prompts for social posts, built on Anthropic's prompting rules, with a fill-in checklist and a verification loop for invented numbers.
By the Superdirector editorial team. Anthropic's documentation pages and three vendor prompt guides, read November 14, 2026.
At a glance
What are the best Claude prompts for social media posts?
The best Claude prompts for social media posts are the ones that state the job completely: the platform, the audience, the voice pattern, the output shape and the one thing the post must do. Anthropic's own prompting reference is the standard to build against: Claude responds well to clear, explicit instructions; examples are its most reliable lever for format, tone and structure; and three to five well-chosen examples make that steering consistent. This page turns that method into eight copy-paste prompts: a Voice Primer and a Platform Set to set context, a Rebuild Prompt, Hook Batch, Carousel Skeleton and Comment Reply Bench to make posts, an Assumption Audit to stress-test an idea, and a Weekly Fill to route a week of slots. Fill every bracket, check every draft against the checklist, and verify every number before anything posts.
- A strong prompt names role, voice, rules and output shape in full; vague topic-and-tone prompts are where generic posts come from.
- Three to five of your own example posts, pasted in tags, are Anthropic's tested lever for voice and structure.
- Prompts make drafts, not outcomes: check every draft, verify every number, and take platform limits from the platform's own pages.
The four layers this pack's prompts share, drawn from Anthropic's rules on role, examples and numbered steps. Original diagram; source basis: Anthropic's prompting reference and the prompt templates in this article.
Make it work for your brand
Find it. Remake it. Post it. Superdirector turns the viral formats in your niche into your brand's carousels, X posts and video scripts, and keeps the calendar full.
The short answer: eight prompts to start with
The prompts that keep working across platforms do three things a vibe-check prompt never does: they name who is writing, they carry the rules Claude cannot guess, and they fix the shape of the answer. Most Claude prompts floating around social media name a topic and a tone and stop there. That under-specifies the job, and generic output is the honest result.
This pack has eight. Two set context (the Voice Primer and the Platform Set). Four make posts (the Rebuild Prompt, the Hook Batch, the Carousel Skeleton, the Comment Reply Bench). One stress-tests an idea (the Assumption Audit). One routes a week (the Weekly Fill). Each is copy-paste ready, with slots in [BRACKETS] you fill before running. The pack sheet download in this section holds all eight in one CSV.
- Voice Primer: turn three to five of your posts into a reusable voice card.
- Platform Set: one idea, native versions for X, Instagram, LinkedIn and TikTok.
- Rebuild Prompt: turn long-form you already wrote into two social posts.
- Hook Batch: five opening lines in three named styles, then pick one.
- Carousel Skeleton: a slide-by-slide outline, one point per slide.
- Comment Reply Bench: one-line replies in your tone, spam marked.
- Assumption Audit: the three assumptions in a draft most likely to be wrong.
- Weekly Fill: route your stated slots without writing final copy.
Two honest notes before the pack. First, prompts make drafts; nobody has published a tested comparison of prompt packs against real account results, so read the checking section before anything posts. Second, the prompts encode platform shape rules as editable craft defaults, and platform hard limits come from the platform's own pages, which section four covers.
What a strong Claude social prompt contains
Anthropic's official prompting reference is the anchor for everything in this pack, and its rules are short. Claude responds well to clear, explicit instructions. Write the output you want explicitly, because if you want more than the literal task, you have to ask for it. When order or completeness matters, Anthropic says to provide instructions as numbered steps.
Anthropic's own image for the job is a new hire: Think of Claude as a brilliant but new employee who lacks context on your norms and workflows. Its key test belongs in your head every time you trim a prompt: show it to a colleague with minimal context on the task, and if they would be confused, Claude will be too. Three structure rules finish the setup. A one-line role in the system slot focuses behavior and tone. Examples steer format, tone and structure, and Anthropic's guidance says three to five wrapped in tags work best, tagged so Claude can tell examples apart from instructions. With long inputs, place the documents at the top of the prompt and write your ask after them; Anthropic measured up to a 30 percent response-quality improvement from ending queries that way in its own tests.
Why bother with all this rigor for a social post? Anthropic's business write-up attributes to prompt engineering real numbers it reported from one Fortune 500 deployment: we improved Claude’s accuracy by 20% - and helped get their product to market faster and at a lower cost. That is Anthropic's claim about its own engagement, not a promise for your posts. The same overview carries the counterweight worth quoting to yourself: Not every success criterion or failing eval is best solved by prompt engineering. If a prompt keeps failing after real editing, the fix may be a different tool or a different job, not a longer prompt.
One date stamp of honesty: Anthropic's reference notes techniques measured on one model should be re-checked on yours. Run the pack's outputs against your own eyes each time a model changes, and keep dated copies of the prompts that worked.
The eight prompts
Every prompt below runs in plain Claude chat, with no Claude Code, no API key and no setup. Fill the [BRACKETS] first. If a slot is empty when you paste, the prompt is built to ask you rather than guess.
1. The Voice Primer. Run this once, save the card it returns, and paste that card into every other prompt in the pack.
You are a brand voice analyst.
<example 1>
[PASTE ONE POST THAT SOUNDS EXACTLY LIKE YOU]
</example>
<example 2>
[PASTE ONE POST THAT SOUNDS EXACTLY LIKE YOU]
</example>
<example 3>
[PASTE ONE POST THAT SOUNDS EXACTLY LIKE YOU]
</example>
Read the examples and return a voice card with:
1. Sentence length pattern and rhythm
2. How posts open, and the opening move that appears most
3. How posts close
4. Five habits to keep and three habits I never use
5. Words I would never say
End with one test paragraph about [TEST TOPIC] in that exact voice, under 80 words. If any slot above is empty, ask me for it before writing anything.2. The Platform Set. One idea in, native shapes out. The platform blocks are craft defaults you can edit; hard limits come from the platform's own pages.
You are my social media editor. I do the thinking; you draft the versions.
<voice>
[PASTE YOUR VOICE CARD FROM THE VOICE PRIMER]
</voice>
<idea>
[THE CORE IDEA IN ONE OR TWO SENTENCES]
</idea>
Write two versions for each platform. Same idea in every version, its own shape in each.
X: one post inside the limit on X's own help page. The first line carries the whole point. At most two hashtags.
Instagram: a caption whose first line hooks before the fold, middle delivers value, last line holds one call to action.
LinkedIn: a short opening line, then three short paragraphs that state the practical part plainly, ending with one question to readers.
TikTok: a spoken opener for the first three seconds, the point in plain talk, then one line that invites a reply.
Ask me before writing if any slot above is empty. Never invent numbers, quotes or results.3. The Rebuild Prompt. This one encodes Anthropic's long-context rule on purpose: your full text sits at the top, and the ask sits at the end, after the source.
[PASTE YOUR FULL LONG-FORM TEXT HERE, BEFORE ANY INSTRUCTION]
Now, from the text above, produce two posts.
1. An X post that states the single most useful point, inside X's own posted limit, no invented figures.
2. An Instagram caption that retells the same point for a reader who will not open the article.
Mark anything you paraphrase rather than quote with [PARAPHRASE] so I can check it.
<voice>
[PASTE YOUR VOICE CARD]
</voice>4. The Hook Batch. Five openings, three named styles, then you pick before any full post gets written.
Here is my post draft: <draft>
[PASTE DRAFT]
</draft>
Write five opening lines for it. No two may carry the same idea.
Lines 1 and 2: contrarian. Name the common habit, then the correction.
Lines 3 and 4: specific outcome. Name who gets what practical result, with no invented figures.
Line 5: a plain statement worth a screenshot.
<voice>
[PASTE YOUR VOICE CARD]
</voice>
Then stop and ask me which line to keep before writing the full post.5. The Carousel Skeleton. One point per slide, one line per headline, and the deck can be any number of slides the idea actually needs.
Build a carousel outline for [TOPIC] aimed at [AUDIENCE].
Rules:
- Slide 1 is the hook: one line that sets the stakes without hype.
- One point per slide; split a slide that needs two.
- Every headline fits on one line.
- The last slide holds exactly one call to action, then stops.
Return a table with columns Slide, Headline, Point. Under the table, one line: what I should cut first if the deck feels long.6. The Comment Reply Bench. Replies come back one line each, with spam separated so you are not writing to bots.
Here are comments on my post. Draft one-line replies in my tone: <tone>
[ONE LINE OF TONE, FOR EXAMPLE: DIRECT, WARM, NO FILLER]
</tone>
<comments>
1. [PASTE COMMENT]
2. [PASTE COMMENT]
3. [PASTE COMMENT]
</comments>
Rules: one line per reply; each reply adds something instead of agreeing louder; mark marketing spam SKIP SPAM; flag any comment I should answer myself; never invent facts about my product to satisfy a comment.7. The Assumption Audit. Asking a model to defend its own draft beats hoping it volunteers its doubts; the useful instruction asks which assumptions are most likely wrong, which is the move one vendor guide credits for the honest answers in its own testing.
Here is a draft post: <draft>
[PASTE DRAFT]
</draft>
Before I post it, list:
1. The three assumptions inside this draft most likely to be wrong, one line each.
2. What observation would disprove each one.
3. Which sentence should change if an assumption fails, and the safer sentence to put in its place.
Do not rewrite the whole draft. If a claim is false, say so plainly; do not soften it.8. The Weekly Fill. A routing pass, not a writing pass: it matches your stated slots to the prompt you should run, so final copy is written where it belongs, in the platform prompt.
My inputs:
- Topic list: [3 TO 6 TOPICS]
- Slots: [FOR EXAMPLE: X DAILY, INSTAGRAM TUESDAY AND FRIDAY]
- Standing events: [A LAUNCH, A HOLIDAY, A RECURRENCE]
Draft the stated slots only. For each slot give: the slot, the platform, which pack prompt to run, and one line on the angle. Leave unplanned slots empty with a one-word note. Do not write final copy in this table.Which prompt when: a short routing table before the next sections.
| Your situation | Start with | Then |
|---|---|---|
| A whole account, nothing posted yet | Voice Primer, then Platform Set | Weekly Fill to plan the slots |
| A long article nobody read | Rebuild Prompt | Hook Batch for the same idea |
| The same caption pasted everywhere | Platform Set on one idea | Carousel Skeleton when it earns slides |
| Comments piling up faster than answers | Comment Reply Bench | Assumption Audit on the reply you doubt |
| A draft you do not fully trust | Assumption Audit | Hook Batch after the fix |
The same prompt in two orders: Anthropic reports up to a thirty percent response-quality gain when the long input comes first and the ask ends the prompt. Original diagram; source basis: Anthropic's long-context guidance quoted in this article.
Adapt a prompt before you reuse it
Anthropic's example rules apply to your slots, not only to Claude's output. An example should mirror the actual use case closely, and your examples should cover the range you need without teaching an accidental pattern. When you paste example posts into any prompt in this pack, keep them wrapped in tags so Claude can tell examples from instructions, exactly as the official guidance describes.
Platform rules deserve one careful sentence. The blocks inside the Platform Set (first line placement, caption shape, a two-hashtag ceiling) are craft defaults this pack chose, meant to be edited. The hard facts, a character cap or where a caption truncates, belong to the platform, and the vendor guides read for this article state versions of those mechanics without platform sources. So do this: take each platform's current limits from the platform's own help pages, paste them into your platform block, and date the line in your prompt file. A rule with a date on it is a rule you can trust; a rule without one is a rumor in six months.
Same for the audience slot. Weak fill in, weak output out. "My audience is founders" will draft like a billboard. "My audience is solo B2B founders under ten employees who post weekly and get no replies" will draft like a person. That trade is the single highest-leverage minute in this pack, and the colleague key test from section two double-checks it: if a stranger reading your finished prompt would ask "which audience?", the slot is not filled.
Check every draft before it posts
A draft is a question, not a decision. The checking loop below is short enough to run every time, and it is the difference between a prompt pack and a mistake factory.
- Read it against your voice card. The card is the standard; a draft that sounds like a press release fails it.
- Count the calls to action. One per post. Two asks in one post is no ask.
- Verify every statistic, date and quote against a primary source, and cut what you cannot verify. The vendor guides this article drew on repeat the same warning from their own corners: treat a model's confident figure as unverified until you check it.
- Keep customer data, internal strategy and credentials out of every prompt. Placeholders, not details.
- Take platform mechanics from the platform's own pages, dated.
Let Claude check itself before you read. Anthropic names self-correction, a draft, then a review against criteria, then a refine, as its most common chaining pattern, and that loop works in a single chat: run the draft prompt, paste your checklist, and ask Claude to score its own draft against each line before you read a word. The Assumption Audit does the same job for the claim inside the post.
What the loop is not: it is not a guarantee, and it is not autopilot. The vendor guides sell on phrases like "Claude drafts. You edit." That line is correct and it is the honest center of all of this: the prompts compress the blank page, and the judgment in front of the publish button stays yours.
The checking loop from section five, run on every draft before anything posts. Original diagram; source basis: Anthropic's draft, review and refine chaining pattern and this article's checklist.
Teach Claude your brand voice with three to five posts
The Voice Primer is the prompt that pays for the whole pack, and the official rule behind it is short: examples are the most reliable lever, and three to five well-picked posts carry it. Pick your best performers, varied: one explainer, one plain statement, one story if you write stories. Pasting ten near-identical captions teaches rhythm, not range, and the model will inherit the sameness.
The card is disposable in the best way. Save the text it returns in your prompt file, paste it into every other prompt, and regenerate it when the brand changes: new positioning, a rename, a model change, or a month where the posts drifted. Anthropic's caution about model-specific techniques applies here more than anywhere; the same card drafts noticeably differently across model generations, so date your card with the model you used.
What the card is not: it is not a promise of engagement, and it is not proof a post sounds like you. The reader of the test paragraph is you. If the card's test paragraph fails your own ear, redo the primer with different example posts before blaming any model.
Stop invented numbers before they post
Every prompt in this pack ends with the same instruction, and it is not decoration: never invent numbers, quotes or results. Claude drafts plausible specifics with confident tone, which is exactly the failure mode a social post cannot survive, and no model version has fixed it by itself. The loop below closes it in four steps.
- State the ban in the prompt, as this pack does, and ask Claude to mark paraphrases so nothing slips in as an unmarked borrow.
- Read the draft hunting specifics: a figure, a date, a named study, a customer line. List every one.
- Verify against a primary source or cut. A post without the number beats a post with a wrong number, every time.
- Rerun with sources pasted in. When you do have a real figure, paste the source at the top of the prompt and ask Claude to ground the line in it, the way Anthropic's own case example does with quoted source text before the answer.
One more official technique worth borrowing for anything hard: in its documented deployment case, Anthropic had Claude show its work in a hidden scratchpad before the customer-visible answer, a step nobody sees that measurably steadied the final text. In plain chat, the same move is one sentence: "work out the specifics on a scratchpad first, show me only the checked final draft." Then the Assumption Audit asks the question the model will not volunteer: which three assumptions in this draft are most likely wrong?
Hands-on or handed off: keep it or delegate it
There are two honest answers to "what is the best prompt," and they split by who is working. If you are working, this pack is it, and the time it asks of you is real: fill slots, read drafts, check numbers. If the work itself is what you want off your desk, prompts are the wrong tool no matter how good they are, because a prompt only helps someone who is sitting there to run it.
Superdirector publishes this guide, so weigh what follows accordingly. On Starter at $29 a month, the plan is built for this exact hands-on job: work from Claude or ChatGPT with live social data, ideas, scripts and downloads, 300 credits a month, and a 7-day free trial with a card required. On Growth at $199 a month, the shape changes: scheduled research, routines and supported Instagram and X publishing on your accounts once you turn posting on, with the editorial calls staying with you. Finished AI video and fully autonomous channels are marked upcoming on the pricing page, so confirm current scope and availability there before deciding.
Simplest decision line: if the bottleneck is drafts, run the pack; if the bottleneck is that no one is sitting there running anything, delegate the account and keep the editorial calls. Both are fine jobs; they are just different ones.
The two honest routes for the same job, split by who is doing the work. Original diagram; source basis: the plan details and upcoming labels on Superdirector's pricing page.
Your first half hour with the pack
A first session that ends with two checked posts, in about thirty minutes:
- Minute 0: write one sentence of success criteria before any prompt runs. What does a good post look like, for one platform, in one sentence? Anthropic's overview starts every prompting job exactly there, and the metaprompt recipe in its cookbook is the documented route when no first draft exists yet.
- Minutes 1 to 8: run the Voice Primer on three of your best posts. Read the card, fix it, save it dated.
- Minutes 9 to 18: run the Platform Set on one real idea. Read both platform versions.
- Minutes 19 to 27: run the checklist on each version, then the Assumption Audit on the one you will actually post, and make the fixes.
- Minutes 28 to 30: save the prompts that produced something usable to the pack sheet, with the date and the model you used, so next week starts from a prompt that already worked.
The recurring cost after that is minutes, not sessions: a rerun of the Voice Primer when the brand moves, the checklist every time, and a dated prompt file that gets stronger every week you work in it.