Ali-Aria/optimize-amazon-title-highlights

Optimize Amazon Titles and Item Highlights

Create conservative recommendations only; never modify Seller Central. Read references/amazon-policy.md for policy decisions and references/input-output-schema.md for fields. Use scripts/marketplaces.mjs as the marketplace source of truth.

Choose the mode

• Direct: one URL, screenshot, page text, ASIN, existing title, or product record. Return copy and validation in chat.
• Batch: multiple records or CSV/XLSX, up to 100 rows. Produce the review workbook.

Process

1. Extract only supported facts: marketplace, exact brand, product type, model, compatibility, variant attributes, materials, features, and use cases. A screenshot supplies visible facts only.
2. Set productClass to media, non_media, or unknown. Only explicit media is exempt from 75/125; unknown uses the limits with an editorial warning.
3. Write in the marketplace language. Keep brand, model, compatibility, measurements, counts, and units exact. Put identity and key variants first; move secondary verified details to Item Highlights.
4. Do not mechanically truncate, invent claims, stuff keywords, repeat the title, or fill unused characters.
5. Validate the exact strings with the bundled scripts. Revise hard errors; retain editorial warnings for review.

Hard limits apply only when not Media: title ≤75, Item Highlights ≤125. There is no minimum or target length.

Validate

Direct mode:

node scripts/validate_copy.mjs single-record.json

Return:

推荐标题（N/75）：
...
Item Highlights（N/125）：
...
硬校验：通过 / 需修改
编辑建议：无 / ...
身份校验：通过 / 需修改
改写说明：...

Batch mode:

node scripts/normalize_input.mjs input.txt|input.csv|input.xlsx records.json
node scripts/validate_records.mjs records.json validated.json
node scripts/build_review_workbook.mjs validated.json output.xlsx preview-directory

For batch work, copy the bundled scripts into a writable task directory. Follow the installed Spreadsheets skill to link that directory's node_modules to its provided packages (symlink on macOS/Linux; junction or symlink on Windows), then run the commands there. Never bundle or install dependencies yourself. Deliver only the final workbook. Preserve duplicate/error rows, evidence URLs, retrieval time, confidence, hard errors, editorial warnings, and review status.

If evidence is insufficient or conflicting, use 资料不足, leave proposed copy blank, and state the missing facts. Treat every result as a human-review recommendation.