TL;DR
You have one asset and five minutes. Don’t copy a single caption to every platform — a caption that reads native on X dies on LinkedIn, and Xiaohongshu has its own grammar. Feed AI the asset, the per-platform reader, your banned tics, and 2-3 recent reference captions you admire, then ask for 3 tonal variants per platform (punchy / curious / contrarian). The biggest 2026 change: Instagram enforces a hard 5-hashtag cap (rolled out December 2025), so any “use 10-30 tags” prompt is now wrong.
The task
You have an image or short video and you want it live across X, LinkedIn, Instagram, and Xiaohongshu in one pass. The failure mode is writing one caption and pasting it everywhere. The job is to use AI for three variants per platform, each adapted to that platform’s reader, length norm, and CTA verb — in a tone that doesn’t announce itself as AI.
Per-platform rules (as of June 2026)
Hardcode these into your prompt so the model stops guessing. The hashtag column is where most stale prompts go wrong.
| Platform | Preview before “more” | Hard caption limit | Hashtags | CTA verb that lands |
|---|---|---|---|---|
| X (Twitter) | ~280 chars (full post visible) | 280 free / 25,000 on Premium | 1-2 max, one word each | Reply, Quote, Bookmark |
| ~140 chars (first 3 lines) | 3,000 (sweet spot 1,300-1,900) | 3 niche | Comment, DM, share | |
| ~125 chars | 2,200 | 5 max (enforced) | Save, Share to story | |
| Xiaohongshu (小红书) | title ~20 chars | body ~1,000 chars | 3-5, mix broad + niche | Collect (收藏), Follow |
Two practical consequences. First, front-load: X shows the whole post, but LinkedIn truncates after ~140 characters and Instagram after ~125, so the hook and the CTA both belong in the first line. Second, Instagram’s old 30-tag habit is dead — as of December 2025 the app caps you at 5, and Adam Mosseri has said a few specific tags beat a long generic list, so a tight prompt should ask for exactly 5 niche tags, not “5 to 10.”
When AI helps — and when it does not
AI is strong at varying tone (punchy / curious / contrarian) and switching vocabulary register per platform. It is weak at current native idiom: what reads as cringe on TikTok this month was often fine last quarter, and the model’s training lags. Feed it 2-3 recent captions you admire as reference; without them it defaults to a stale “creator” voice (the “Let’s dive in” / ”🎯” register). Any frontier model handles this — GPT-5.5, Claude Sonnet 4.6, or Gemini 3.1 Pro all produce usable variants — so pick on price and habit, not capability. The input pack matters far more than the model.
What to feed the AI
- Asset description (image / video / carousel: what is visible, what is happening)
- Platform list with the target reader per platform
- Brand voice and banned tics (no ”🎯,” no “Let’s dive in,” no emoji bullets)
- CTA per platform — each rewards a different ask (see table above)
- One recent reference caption you admire, per platform
- Hashtag policy per platform: Instagram exactly 5, X 1-2, LinkedIn 3, Xiaohongshu 3-5
Copy-ready prompt
Write 3 caption variants per platform.
Asset: [description of what's in the image / video]
Brand voice: [one sentence]
Banned tics: [list, including "🎯", "Let's dive in", emoji bullets]
For each platform, here's the reader and a reference caption I admire:
X (Twitter)
Reader: [line]
CTA: [reply / quote / no CTA]
Reference: "[paste]"
Hashtags: 1-2 max, single word each
Length: front-load; whole post is visible
LinkedIn
Reader: [line]
CTA: [comment / DM / share]
Reference: "[paste]"
Hashtags: 3 niche
Length: hook in first 140 chars (rest is hidden behind "more")
Instagram
Reader: [line]
CTA: [save / share to story]
Reference: "[paste]"
Hashtags: exactly 5 niche (Instagram caps at 5 since Dec 2025)
Length: key line in first 125 chars
Xiaohongshu (小红书)
Reader: [line]
CTA: [collect / follow]
Reference: "[paste]"
Hashtags: 3-5, mix of one broad and several niche
Length: title under 20 chars, body conversational
Return per platform: 3 variants in these tones — punchy / curious / contrarian.
For each variant include: caption text, hashtags, and one sentence on which
type of reader is most likely to engage.
Do not reuse the same opening word across variants. Vary length within each
platform's norm.To repurpose the same asset into video: add “Now produce a 12-second voiceover script the same asset could carry on TikTok / Shorts / Reels.”
How to check the output is usable
- Each variant matches its platform’s reference vocabulary, not a generic register
- Hashtag counts respect the table — especially Instagram at 5, not more
- The hook and CTA sit inside the visible preview (first ~140 chars on LinkedIn, ~125 on Instagram)
- CTAs use the native verb (Reply on X, Comment on LinkedIn, Save on Instagram, Collect on Xiaohongshu)
- Opening words differ across variants — no repeated “AI cadence”
- A stranger from your target audience would re-share the strongest variant
Common mistakes
- One caption across all platforms: it underperforms everywhere
- LinkedIn voice on X (over-formal) or X voice on LinkedIn (too curt)
- Hashtag overstuffing on Instagram — the 30-tag habit is gone; the app now hard-caps at 5
- Burying the CTA past the “more” cutoff on LinkedIn and Instagram
- Letting AI reuse its favourite emoji: they age fast
- Captions that summarise the asset; a caption should add, not narrate
FAQ
- Should I A/B captions on the same asset? Yes, by reposting at different times rather than firing identical copies. Platforms deprioritise identical re-posts within 24h.
- What about alt text? Always write it. It is an accessibility requirement, and platforms also parse it for discovery.
- Should AI write the hashtags? Yes, but verify each niche tag actually exists and is active — models invent plausible-looking tags that have zero posts.
- Is Instagram really limited to 5 hashtags now? Yes. The cap rolled out across posts and Reels starting December 2025 and is enforced app-side, so prompts asking for 10-30 tags are out of date.
- Which AI model is best for captions? Capability is not the bottleneck here — GPT-5.5, Claude Sonnet 4.6, and Gemini 3.1 Pro all do this well. Your reference captions and banned-tics list decide quality far more than the model choice.
Related
- X thread AI: when the caption becomes a thread
- Cross-platform repurpose: multi-platform reuse
- Content calendar: when captions ship
- TikTok caption: TikTok-specific patterns
- Reel hook: when the asset is a Reel and needs a hook
- Reel caption prompts: additional Reel caption phrasings
- TikTok caption prompts: additional TikTok caption phrasings