follow-amazon-daily

Generate a daily Amazon seller intelligence digest. Default language is Chinese.

This skill is split into two stages, like follow-builders:

1. scripts/prepare-digest.js fetches sources deterministically and prints a single JSON blob (raw items + prompts + config). It does not write the final digest.
2. You, the agent, read that JSON and remix it into a real editorial digest following prompts/daily-digest.md and prompts/translate.md. You write digest/<date>.md and optionally deliver it via scripts/deliver.js.

The point of the split: summaries and translation are written by you from the raw excerpt/body, never by a hardcoded phrase dictionary. Every item gets its own specific summary.

Hard Rules

1. Public sources are the default. Authenticated WeAreSellers and Billion Dollar Sellers content is optional and only enabled when WEARESELLERS_COOKIE or BDS_COOKIE is present.
2. Never ask for or store account passwords. Only use cookie/session env vars or a local logged-in browser workflow the user adds later.
3. Authenticated full text is ephemeral. Use it to derive short insights only. Never write raw HTML, raw article text, or full subscriber/community posts to data/feed-amazon.json or digest/<date>.md. privateItems are stdout-only.
4. WeAreSellers is a community pain-signal source, not policy authority. Confirm policy claims with official sources.
5. Billion Dollar Sellers is industry opinion unless confirmed elsewhere.
6. No auto-publishing to social platforms (Notion, Xiaohongshu, LinkedIn, X, Instagram). Delivery is limited to what the user explicitly configures in config.delivery (stdout default, or their own Telegram / Email / Feishu). Never ask the user to paste a secret in chat — write it to their env file.
7. Follow config.language exactly. See prompts/translate.md.
8. Never fabricate or speculate. Every digest item must carry its original link; no link → exclude it. See prompts/daily-digest.md.

Outputs

data/feed-amazon.json   # deterministic raw public feed (script-written)
digest/YYYY-MM-DD.md     # editorial digest (agent-written, no private items)
stdout                   # full JSON blob from the script (for the agent)
state-feed.json          # per-install cross-run dedup state — LOCAL ONLY

state-feed.json is gitignored runtime state. Never commit it and never ship it: a populated state would pre-mark every item "seen" and make a fresh user's first digest empty. A clean install has no state file, so the first run shows everything (the welcome digest works).

First Run — Onboarding

Gate: if config/sources.json does not have onboardingComplete: true, run this flow before anything else. Drive it conversationally — you edit the config and env files for the user; never make them hand-edit JSON, and never ask them to paste a secret into the chat.

Step 1 — Intro

One or two lines: this is a daily Amazon-seller intelligence digest covering Official / Policy, Seller Ops, Community Pain Signals, Podcast / Video Playbooks, and Newsletter / Analyst Signals. Read config/sources.json and tell them the actual source names and count. One line on the two-stage model (a script fetches raw content; you remix it into a real digest).

Step 2 — Language

Ask: Chinese (default), English, or bilingual → set config.language.

Step 3 — Frequency, time, timezone

Ask cadence: daily (default) or weekly; the local time (HH:MM); their IANA timezone (e.g. America/Los_Angeles, Asia/Shanghai); and the weekday if weekly. Set frequency, deliveryTime, timezone, weeklyDay.

Step 4 — Delivery method

Present four options. For every keyed option, the secret rule is absolute: the user gets the secret from the provider, and you write it into their env file (~/.follow-amazon-daily.env, which the cron sources) — they never paste it into chat, and you never echo it back.

• In-chat / stdout (default) — just show the digest here. No keys.
• Telegram — guide: open @BotFather → /newbot → copy the bot token; send the new bot any message; then curl -s "https://api.telegram.org/bot<TOKEN>/getUpdates" to read chat.id. You put TELEGRAM_BOT_TOKEN in the env file and delivery.chatId in config.
• Email — sign up at resend.com (free), create an API key. You put RESEND_API_KEY in the env file and the recipient in delivery.email.
• Feishu (飞书 / Lark) — in the target Feishu group: 设置 → 群机器人 → 添加机器人 → 自定义机器人 → copy the Webhook URL. Optional: enable 签名校验 and copy the secret. You put FEISHU_WEBHOOK (and FEISHU_WEBHOOK_SECRET if signed) in the env file. The target is whichever group the bot is in — no chatId needed. Set delivery.method: "feishu".

When writing the env file, create/append ~/.follow-amazon-daily.env with export KEY=value lines and tell the user the file path only.

Step 5 — Save config

Write all choices into config/sources.json (language, frequency, deliveryTime, timezone, weeklyDay, delivery.method + chatId/email as needed) and set onboardingComplete: true. Show a short plain-language summary of what you saved.

Step 6 — Schedule (system crontab)

Skip for stdout/on-demand. For telegram/email/feishu, install a crontab entry that runs scripts/run-daily.sh (it does prepare → agent remix → deliver, with a labelled raw fallback if the Claude CLI is unavailable on the cron run):

CRON_TZ=<their IANA tz>
<min> <hour> * * <*|weekday>  /abs/path/to/skill/scripts/run-daily.sh

If CRON_TZ is unsupported on their cron, convert their local time to UTC for the schedule instead. Then verify by running scripts/run-daily.sh once and confirming the message actually arrived in their channel. Be honest about the raw-fallback caveat (no agent in the cron environment = structured feed, clearly labelled, not the full editorial digest).

Step 7 — Welcome digest

Immediately run the full pipeline once now (prepare-digest → you remix per the prompts → deliver via their chosen method → node scripts/mark-seen.js so the first scheduled run won't repeat today's items) so they see real output today. Then ask for feedback — length, focus, tone — and apply it (edit the relevant prompts/ file or config), and confirm what changed.

Step 8 — Reconfigure-by-chat reminder

Tell them everything is changeable by just asking: "switch to English / make it weekly / change time to 8am ET / send it to Feishu instead / make summaries shorter / show my settings". You edit config or prompts for them — and when the time/timezone/frequency changes, you also update the crontab entry.

Daily Workflow

From the repo root:

node scripts/prepare-digest.js          # prints JSON blob to stdout

Then:

1. Parse the JSON blob (config, items, privateItems, prompts, stats, errors).
2. If stats.publicItems is 0, tell the user no public signal was captured and stop.
3. Remix into the digest following blob.prompts.dailyDigest. Translate per blob.prompts.translate and blob.config.language. Base each summary on the item body first, then excerpt. Never reuse a sentence across items.
4. Write the public digest to digest/<date>.md (must contain no privateItems content).
5. Deliver: if config.delivery.method is telegram / email / feishu, pipe the digest to node scripts/deliver.js --file digest/<date>.md. Otherwise (stdout) show it here.
6. Only after the digest is written and delivered, run node scripts/mark-seen.js so the next run excludes these items. prepare-digest.js filters against dedup state read-only and never persists it — if you skip this step (or the run fails) the items simply reappear next time rather than being silently lost. Skip mark-seen if you ran with --ignore-state.

For scheduled runs this whole flow (including mark-seen, only on