TL;DR
One product, three jobs: a detail page (informational), hero one-liners (interruptive), and a short-video voiceover (conversational). Pasting one block into all three is why nothing converts. Below are three copy-paste prompts plus the real platform limits each format has to live inside (Amazon: 5 bullets at 255 characters each, no emoji; Shopify meta description 150-160 characters; TikTok caption truncates around 100-150 characters even though the field holds 4,000). For brand-voice consistency across all three, Claude Sonnet 4.6 needs the least editing; use GPT-5.5 when you want a wide spread of hook variants to pick from.
The task
You launched a new SKU last week — a 30-second fast-charge phone bank. You need detail-page copy for Amazon and Shopify by Wednesday, three hero-image one-liners for the Meta ads going up Friday, and a 30-second voiceover script for the TikTok creator who agreed to post next week. Same product, three completely different formats, three different reading contexts, three different jobs. The intern wrote one block and pasted it into all three slots. The detail-page reads fine, the hero lines are unreadable at 1080×1080, and the voiceover sounds like someone reading a spec sheet aloud.
Where AI helps — and where it does not
AI is genuinely good at producing the three formats in parallel and keeping voice consistent across them — same brand, same posture, three different jobs. It is also good at killing “revolutionary / game-changing / must-have” once you put those on the banned list and feed it 2-3 real lines from past products you liked. Where AI fails: picking which pain point leads the detail page. That decision belongs to your conversion data and your real customer reviews, not to a model that has not read either. Feed AI your top-3 pain points pulled from actual reviews, not from your imagination — otherwise it invents pains that sound plausible and convert nothing.
A common failure mode: AI defaults to the same temperature across all three formats. The detail page reads like an ad, the ad reads like a manual, the voiceover reads like a press release. Tell the model explicitly: detail page is informational, hero line is interruptive, voiceover is conversational.
Which model
For customer-facing copy where brand voice has to hold across dozens of products, Claude Sonnet 4.6 (Claude Pro, $20/mo as of June 2026) needs the least cleanup and drifts off a pasted voice anchor the slowest. Step up to Claude Opus 4.7 only when a single high-ticket page is worth the extra cost. GPT-5.5 (ChatGPT Plus, $20/mo) is the better pick when you want raw variety, like generating 20 hero one-liners to A/B test, because it spreads wider before you narrow down. Gemini 3.1 Pro (Google AI Pro, $19.99/mo) is a fine third option if you already live in Workspace. Whatever you use, the prompt does the heavy lifting; the inputs below matter more than the model.
What to feed the AI
- Product name, category, and one-sentence factual description (no marketing words)
- The top 3 pain points pulled from real customer reviews (not from your team’s assumptions)
- The 3-5 concrete features that solve them — feature, not benefit; let the model bridge to benefit
- Target audience in one sentence — age, motivation, what they currently use instead
- Brand voice anchor — paste 2-3 lines from a past product page you were proud of
- Banned words list — your specific allergens (“game-changing,” “revolutionary,” “must-have,” “ultimate,” any emoji unless the brand uses them)
- Required numbers — at least one quantified spec or comparison per format (30 seconds, 5,000 mAh, 40% lighter, etc.)
- Price tier — premium / mainstream / value changes voice register; premium copy is calmer and uses fewer exclamation marks
Platform limits each format lives inside
Tell the model the hard limit up front, or it writes copy that gets truncated or rejected. These are current as of June 2026; check your seller dashboard before launch, because Amazon adjusts category rules quietly. Amazon publishes its bullet rules in Seller Central’s bullet-point requirements.
| Surface | Hard limit | Notes |
|---|---|---|
| Amazon bullets (3P seller) | 5 bullets, 255 chars each | Best practice under 200 for mobile scanning; emoji are prohibited and can suppress the listing |
| Amazon product description | 2,000 chars | HTML tags deprecated since 2021, so no bold or lists in this field |
| Shopify meta description | 150-160 chars | Google truncates past ~160 on desktop, ~120 on mobile |
| Shopify body copy | 80-120 words recommended | Auto-generated if you skip it, which is fine for storage, bad for SEO |
| TikTok caption | Field holds 4,000 chars | Display truncates around 100-150 chars, so front-load the hook in the first two lines |
| Meta / hero image | 6-12 words to stay legible | Must read in ~1 second at 1080×1080 on a phone |
Three formats, three prompts
1. Detail-page structured copy
You are a 5-year e-commerce ops lead writing for {Amazon / Shopify / Tmall / TikTok Shop}.
Write a product detail-page block.
Product: {name, category, 1-sentence factual description}
Real top-3 pain points (from reviews, not assumptions): {paste}
Features that solve them: {list, 3-5}
Target audience: {1 sentence}
Brand voice anchor: {paste 2-3 lines from a past product}
Required numbers: {at least 1 quantified spec must appear}
Banned words: {revolutionary, game-changing, must-have, ultimate, etc.}
Structure:
1) One-sentence positioning. Specific, not aspirational.
2) 3 pain points + 3 matching solutions, each line short.
3) One 80-120 word use-case paragraph that names a specific moment (not "everyday life").
4) Exactly 5 factual spec bullets, each under 200 characters (Amazon caps bullets at 255 and prohibits emoji). No marketing words. Numbers preferred.
5) One gentle CTA. Direct, not pushy.
Voice: useful and warm. If a sentence does not help the buyer decide, delete it.2. Hero-image one-liners (Meta, Shopify hero, Amazon main image)
Give me 10 hero-image one-liners for this product.
Each 6-12 words. At 1080×1080 they must be readable in 1 second.
Product info: {same as above}
Requirements:
1) Line 1: feature + number ("Full charge in 30 seconds").
2) Line 2: replacement framing ("Stop carrying a brick.").
3) Line 3: status / use-case framing ("What pros bring on flights.").
4) At least 4 of the 10 lines include a number.
5) No empty intensifiers ("amazing," "miracle," "must-buy," "you won't believe").
6) No emoji unless the brand voice anchor uses them.
7) Each line must work as standalone — no second-line dependency.
End with: which 3 you would test first on cold audience, with one-sentence reasoning each.3. Short-video voiceover (TikTok, Reels, YouTube Shorts)
You write conversion-driven 30-second voiceover scripts for {TikTok / Reels} creators.
Voice: daily-vlog, not ad. Read it aloud — it should sound like a friend talking.
Product: {info}
Audience: {who watches this creator}
Creator's existing tone: {paste a transcript line from their recent post if possible}
Structure (timestamps matter — voiceover words at 2.5 words/sec):
- 0-3s hook: concrete scene that pulls the target audience in immediately. No "Are you tired of," no question hooks.
- 3-15s: the sharpest pain point + how this product solves it. One specific number.
- 15-25s: one personal-use detail + one comparison ("the one I had before died in 90 minutes").
- 25-30s: CTA. Conversational, not pushy. Name the platform's affordance ("link in bio" not "buy now").
Output:
- The 30-second script with timestamp markers.
- 2 alternative 3-second hooks (so the creator can pick).
- A caption under 150 characters (TikTok truncates the rest behind a "more" tap).
- A 1-line on-screen text overlay for the hook.
Banned: "game-changing," "you won't believe," "this changed my life," emoji unless creator's tone uses them.Shorter variant — single SKU rewrite
Rewrite the {detail page / hero line / voiceover} below.
Current version: {paste}
Problem with current: {what is not working — feels generic / no number / sounds like an ad}.
Return 3 alternatives, each changing exactly one variable (pain point lead, opening hook, or CTA).
Constraint: do not invent specs that are not in the product info above.Sample output
A useful detail-page positioning line: “30-second full charge, fits in a back pocket, 5,000 mAh — the bank you take instead of remembering to charge your phone.” — concrete, has the spec, names the use case.
A useful hero one-liner: “Full charge in 30 seconds. Yes, really.” — number, posture (slight self-awareness), zero adjectives.
A useful TikTok hook (0-3s): “This is the only power bank that has survived 6 months in my bag without becoming garbage.” — specific scene, 6 months as proof, no question hook, no “are you tired of.”
A useful voiceover middle (15-25s): “My last one died at 90 minutes and got thrown out at LAX. This one I have charged maybe a hundred times and it still gets a full charge in under a minute.” — comparison, specific number, real-life moment.
How to refine
- If the detail page reads like an ad: “Cut every sentence that does not help a buyer make a decision. Replace marketing verbs (revolutionize, transform) with the underlying spec or use case.”
- If hero lines are unreadable at thumbnail size: “Test each line at 6 words max. If you cannot read it in 1 second on a phone screen, rewrite shorter.”
- If the voiceover sounds like a press release: “Read each line aloud. If you would not say it to a friend, rewrite. Cut every passive verb.”
- If the three formats sound like three different brands: “All three formats must share the same opening posture. The detail-page positioning, the hero line, and the voiceover hook must sound like the same person wrote them.”
- If the pain points feel invented: “Replace any pain point that does not appear in the customer reviews I gave you. If only 2 are review-sourced, write 2 — do not invent a 3rd.”
Common mistakes
- Same voice across all three formats: detail-page is informational, hero is interruptive, voiceover is conversational; one template across all three is why nothing converts.
- No target audience in the prompt: copy that pulls in no one, because it pulls toward everyone.
- No required numbers: adjective soup, indistinguishable from every other product in the category.
- Letting the model invent specs: “lasts 8 hours” when the real number is 6 destroys trust the moment a buyer notices.
- Forgetting platform constraints: Amazon gives 5 bullets at 255 chars each and bans emoji, Shopify wants a 150-160 char meta description, TikTok truncates the caption around 150 chars; pasting one block into all three is wasted work.
- No brand voice anchor in the prompt: output reads generic, gets archived by your own merch team because it does not match the brand.
- Marketing-speak banned list missing the obvious offenders: “game-changing,” “revolutionary,” “must-have,” “ultimate” — every category leader has stopped using these; banning them in the prompt forces the model to find a real reason to write the sentence.
- Treating short-video voiceover like ad copy: TikTok viewers skip ads in 2 seconds; if your hook sounds like an ad, the algorithm reads dwell time as a fail.
FAQ
- One product, three formats — do I need all three?: Detail page is the floor; it has to exist. Hero lines are required if you run paid ads or update your storefront hero image. Voiceover is only required if you are shipping creator content this cycle. Most brands need 1 + 2; only some need all three.
- Should AI write the headline?: Generate 10 variants and pick by intuition + the read-aloud test. The model is fine at variety but bad at deciding which one is brand-correct. That last call is yours.
- What about voice for high-ticket products?: Calmer, more proof, fewer exclamation marks, no urgency. A $400 product reads like considered purchase; a $20 product can borrow a bit more vibe. The cheap version of premium copy is “premium copy with more exclamation marks.”
- Can the same prompt write copy for B2B SaaS products?: Replace “pain points from reviews” with “pain points from sales call notes,” “hero one-liners” with “above-the-fold landing-page H1s,” and “voiceover” with “30-second product video script.” Structure transfers; vocabulary does not.
- How many real reviews do I need before AI writes detail-page copy?: 15-25 reviews per product is the floor. Below that the pain points are anecdote, not pattern, and the model amplifies the loudest reviewer rather than the most common buyer.
- Which model should I actually use?: For copy that has to stay on-brand across many SKUs, Claude Sonnet 4.6 (Claude Pro, $20/mo) drifts off a pasted voice anchor the slowest and needs the least editing. Use GPT-5.5 (ChatGPT Plus, $20/mo) when you want a wide spread of hook variants to choose from. Reserve Claude Opus 4.7 for one high-ticket page where the extra cost pays back. All figures as of June 2026.