English | 中文

follow-amazon-daily

Daily Amazon seller intelligence digest skill.

It combines the low-friction feed aggregation pattern from follow-builders with the structured, auditable workflow style from wayamzpost. A deterministic script fetches raw content; the agent remixes it into a real editorial digest in Chinese, English, or bilingual mode. Nothing is auto-published to social platforms.

Architecture (two stages)

1. scripts/prepare-digest.js — fetches every source, writes the deterministic raw feed data/feed-amazon.json, and prints a single JSON blob (raw items + prompts + config + stats) to stdout. It does not write the digest and does not translate.
2. The agent — reads that JSON and remixes it into digest/YYYY-MM-DD.md following prompts/daily-digest.md and prompts/translate.md. Every item gets its own specific summary written from the raw body/excerpt — there is no hardcoded phrase dictionary, so the output is not repetitive boilerplate.

What It Produces

• data/feed-amazon.json: deterministic raw public feed (script-written)
• digest/YYYY-MM-DD.md: editorial digest (agent-written, never includes private authenticated content)
• stdout: full JSON blob for the agent, plus optional stdout-only private authenticated signals

Quick Start

npm test
node scripts/audit-sources.js --dry-run
node scripts/prepare-digest.js          # prints the JSON blob
# then the agent remixes it per prompts/daily-digest.md into digest/<date>.md

Language

Default is Chinese (config.language: "zh"). Override at runtime:

node scripts/prepare-digest.js --language en
node scripts/prepare-digest.js --language bilingual

Translation is performed by the agent following prompts/translate.md: natural, native-sounding Chinese, with operator terms and proper nouns kept in English (Amazon, Walmart, TikTok Shop, Sponsored Products, Brand Registry, PPC, ACOS, ASIN, SP-API, Search Query Performance) and URLs unchanged.

Delivery

Default is stdout (show in chat, no keys, no auto-publish). The user can opt into their own Telegram, Email, or Feishu delivery in config/sources.json:

{ "delivery": { "method": "telegram", "chatId": "123456789" } }
{ "delivery": { "method": "feishu" } }

Secrets come from environment variables only (the cron sources them from ~/.follow-amazon-daily.env): TELEGRAM_BOT_TOKEN, RESEND_API_KEY, FEISHU_WEBHOOK (+ optional FEISHU_WEBHOOK_SECRET). Feishu uses a custom-bot group webhook — no app, no chatId. Send a finished digest with:

node scripts/deliver.js --file digest/2026-05-16.md

For an automatic daily run, onboarding installs a system crontab entry that runs scripts/run-daily.sh (prepare → agent remix → deliver, with a clearly labelled raw fallback if the Claude CLI is unavailable on the cron run).

Sources

Defined in config/sources.json (8 sources):

• Amazon SP-API Release Notes (official)
• Amazon Ads Library (official)
• Helium 10 Blog (industry)
• Serious Sellers Podcast — RSS (media)
• AMZ123 跨境头条 — /t link list (industry)
• WeAreSellers — RSS / HTML (community)
• Billion Dollar Sellers — archive (newsletter)
• My Amazon Guy — YouTube RSS (media)

Each source has a lookbackHours window; items older than that, and items already shown in a previous run (state-feed.json, per-install), are filtered out so the digest is only what's new.

Optional Auth

WeAreSellers and Billion Dollar Sellers can be enriched with authenticated cookies:

export WEARESELLERS_COOKIE='name=value; other=value'
export BDS_COOKIE='name=value; other=value'
node scripts/prepare-digest.js

The script never persists raw authenticated page text. Authenticated insights are returned as privateItems in the stdout blob only — never written to data/feed-amazon.json or the public digest.

GitHub Actions

The daily workflow tests the parser, audits sources in non-fatal dry-run mode, refreshes data/feed-amazon.json, and commits only that feed file. The editorial digest is produced by the agent on demand, not in CI. Keep auth cookies out of public workflow logs.

Feed Item Schema

Each item in data/feed-amazon.json:

• id, source, sourceType, category
• title, url, publishedAt
• excerpt (short metadata)
• body (raw readable text for public sources; empty for community / subscriber-gated sources)
• tags, access, sourceReliability
• thin (optional, true when the source returned only an index/JS shell — the agent reports it honestly instead of fabricating a summary)

Seller impact and suggested actions are written by the agent at digest time, not stored as canned strings.

License

MIT