I run e-commerce across Amazon JP, Rakuten, Yahoo, Qoo10, and TikTok Shop Japan. The single biggest bottleneck has always been listings — not products, not logistics, not ads. Writing decent product copy for each SKU, in Japanese, at scale, kills momentum.
Last month I rewired the entire workflow using Claude. This post is the actual playbook, including the prompt I use, where Claude gets Japan wrong, and what the numbers look like.
TL;DR
- Old way: 30–60 minutes per SKU for a usable listing
- New way: ~5 minutes per SKU after a light human edit pass
- Result: roughly [X] SKUs processed in a single day
- The trick is not "ask Claude to write a listing." It's structured inputs + Japan-specific guardrails + a category-level batch pattern.
Why bulk Amazon JP listings are uniquely painful
Anyone who has sold on amazon.co.jp knows the problem isn't translation. It's that Japanese product copy has its own conventions:
- Honorifics shift depending on whether you're describing a gift, a daily-use item, or a luxury product.
- Seasonal vocabulary matters more than in English markets (春限定, 季節の変わり目, etc.).
- Bullet points read very differently — Japanese shoppers expect spec-dense bullets, not the benefit-first style that wins on amazon.com.
- Brand-safe phrasing is narrower. A casual "perfect for everyone" reads cheap in Japanese.
If you machine-translate an English listing, you get something that technically makes sense and emotionally feels off. Conversion drops. You don't see it in keyword tools — you see it in CVR three weeks later.
The new workflow (5 steps)
Step 1: Build a structured input form
Before Claude touches anything, every SKU needs to be reduced to clean inputs. I use a Google Sheet with one row per SKU and these columns:
- Product name (JP and EN if available)
- Category and subcategory
- 3–5 key features (bullet-list style)
- Target keywords (from Cerebro or Helium 10)
- 1–2 competitor ASINs to mirror in tone
- Tone hint: gift / daily / luxury / utility
This is the single highest-leverage step. Garbage in, garbage out is a cliché because it's true. Claude can't recover from a vague spec.
Step 2: Master prompt with Japan-specific guardrails
Here's the prompt structure I land on after about 20 iterations. Adjust the tone block based on the SKU category.
You are writing an Amazon Japan product listing in natural Japanese.
CONTEXT:
- Product: {product_name}
- Category: {category}
- Tone: {tone_hint} // gift / daily / luxury / utility
- Key features: {features}
- Target keywords: {keywords}
- Competitor reference (for tone only, do not copy): {asin}
WRITE:
1. Title (max 200 chars, keyword-front-loaded, Amazon JP conventions)
2. 5 bullet points (spec-dense, not benefit-first)
3. Product description (300-400 chars, neutral-formal tone)
4. 5 backend search terms (no duplicates with title)
JAPAN-SPECIFIC RULES:
- Do NOT use overly polite honorifics (です・ます is fine; avoid いただきます constructions)
- Do NOT translate brand promises literally ("perfect for everyone" → cut)
- Use full-width punctuation (。、)
- Bullets should lead with spec, not adjective
- Avoid katakana-English when a native Japanese word exists, EXCEPT for established
category terms (e.g. "ステンレス" is fine, "クオリティ" usually isn't)
OUTPUT FORMAT:
Plain text. No markdown. No commentary. Just the four sections labeled
TITLE / BULLETS / DESCRIPTION / SEARCH TERMS.
The guardrail block is the part that took the longest to develop, and it's where most teams get this wrong. Without it, Claude produces grammatically perfect Japanese that screams foreign-operated brand to native shoppers.
Step 3: Batch by category, not by SKU
This is counterintuitive. The instinct is to process SKU by SKU, finishing one before starting the next. Don't.
Instead, group all SKUs in the same category and process them in one Claude conversation. Reasons:
- Context carries. After 3 listings in the same category, Claude has internalized the tonal conventions and the rest go faster and more consistently.
- You catch your own input inconsistencies. By the 5th SKU, you notice that your tone hint "daily" produced very different outputs across SKUs — usually because your inputs varied, not Claude's outputs.
- Cheaper. Fewer system prompt reloads.
Step 4: Human edit pass focused on Japan-specific failure modes
I don't edit for grammar. Claude's Japanese grammar is essentially flawless. I edit for these three things:
- Honorific creep. Claude sometimes drifts into 〜していただけます forms that feel salesy in a listing context.
- Generic adjectives. 高品質, 便利, 安心 — these are placeholder words. I replace each with a concrete spec.
- Keyword stuffing artifacts. When the keyword list is long, Claude occasionally repeats a keyword three times in the title. One pass to clean.
This pass takes me about 3–5 minutes per SKU. Compared to writing from scratch, it's a 6–10x speedup.
Step 5: Upload, then iterate based on CVR
Don't try to perfect the listing pre-upload. The data on Amazon JP is the only feedback that matters. I rerun the prompt on underperformers after 30 days using the actual session search terms from the report, fed back into the keyword field.
What Claude Gets Wrong (and Why It's a Feature, Not a Bug)
A few specific failure modes you should expect:
- Over-formal closings. Listings shouldn't end with 「ぜひお試しください」 — sounds like an ad. Strip it.
- Mixing 漢字 and ひらがな inconsistently for the same word across a listing. Quick find-and-replace fixes it.
- Inventing technical specs. If your input is vague, Claude will fill the gap with plausible-sounding nonsense. Always sanity-check numbers, materials, dimensions.
The reason these aren't dealbreakers: they're all catchable in a fast review pass. The hard part — natural Japanese prose at scale — Claude handles. You're left to do the work that requires judgment, not typing.
Why this matters more for Japan than for US/EU
US and EU markets have a deep pool of bilingual copywriters. Japan does not. The supply of professionals who can write good product copy in Japanese — not just correct Japanese — is a real constraint for international sellers.
Tools like Claude meaningfully close that gap, if you guardrail them properly. Without guardrails you get Google Translate with extra steps. With them, you get something that competes with mid-tier human output at 10x speed.
This is the entire premise behind a project I'm building called TOKI Digital Services — productizing this workflow for sellers who want to enter the Japan market without hiring a full-time Japanese copywriter. More on that in a future post.
What I'd Do Differently Next Time
- Build the input sheet template first. I wasted two days iterating prompts when the real bottleneck was input quality.
- Run a 5-SKU pilot per category before committing to batch. Tonal conventions vary more between Amazon JP categories than I expected.
- Save the edited human-finished listings as few-shot examples for future runs. Claude pattern-matches to good examples better than to instructions.
If you're trying something similar
Happy to compare notes. Drop a comment with what category you're working in and what's breaking — I'll respond with what worked or didn't in my version.
— Posting from Osaka. Currently running multi-marketplace EC in Japan and building tooling around it.
Top comments (0)