openclaudia WeChat Moments digest
Turns a WeChat Moments feed into a short ranked digest: what deserves a reply, what is genuinely informative, and what is marketing wearing a personal voice.
The scripts live in a separate repository, not here, because they depend on the WeChat desktop client rather than any marketing API:
https://github.com/OpenClaudia/wechat-moments
git clone https://github.com/OpenClaudia/wechat-moments.git
Read that repo's README before running anything. It documents the setup, the limits, and the risks — this skill is the ranking method, not the plumbing.
Before you touch it
- It does not decrypt anything and never handles keys. It reads a plaintext database that a separate WeChat export tool has already produced. If the user has no export yet, that is step one and it is not this skill's job — the repo lists existing tools for it.
- Moments do not sync in the background. WeChat fetches the timeline only while the Moments window is open. Without a sync the export contains only what the user already scrolled past, so the digest is a replay of what they have already seen. Say so rather than presenting stale posts as today's.
- Tell the user the limits before setting this up on a schedule. It needs WeChat running and logged in, it only stays complete while the machine is awake (sleep gaps are permanently unrecoverable), Tencent may break or patch it at any time, automating the client may be noticed and may carry account risk, and nothing is guaranteed.
- Never post, like, comment, or message on the user's behalf. Reading and writing files is the whole remit.
Ranking
This is the part worth carrying to any feed, not just WeChat.
Score = event weight × relationship weight. Do it yourself from the structured records; do not write a scoring script. The point is judgement — "her mother is in hospital" outranks "went to Kyoto".
Event weight (0-10). Birth, death, serious illness, wedding, divorce, emigration at the top. New job, funding, graduation, a home, a pregnancy just below. Travel and opinions in the middle. Food, gym, weather, scenery at the bottom. Promotion and recruiting score zero.
Insight vs. announcement decides most of the digest, because feeds full of founders and creators are full of engagement bait. Real information has at least one of: a number you could check, a mechanism (why it works), a cost the author actually paid, or a changed mind. Adjectives (突破 / 颠覆 / groundbreaking), a milestone with no mechanism, an invitation, an upvote request, or a takeaway generic enough to predate the event are announcements. Prestige is not evidence — a famous venue is still promotion when the payload is one generic sentence. A tiny post can be real insight.
Relationship weight — lead with how much the user SENT. Not total volume: a marketer who broadcasts dozens of pitches and never gets a reply scores identically to a close friend. A reply is a choice; an inbound blast is not. Never-messaged contacts get filtered hard whatever they posted — expect them to be about a third of any day's feed.
A close friend's small post usually beats a stranger's big news, because only one of them is a relationship the user can act on. Ranking on content alone inverts this every time.
Time since last contact is its own signal. Many messages sent but silent for a long time is a lapsed close friend, and a life event from one of those is the most actionable thing a digest can surface.
Promotion wears a personal voice. A discount code, a signup link, an invite-only pitch: the poster is selling, however casual the phrasing.
Urgency is a modifier, never a reason. Never open a digest with a dated section — that structure promotes whatever has a deadline, and deadlines are what promotions have. An expiring ad is still an ad.
Output
Ten items is plenty; fewer is often right, and zero worth reporting is a legitimate result — say so rather than padding. Keep strong ties and weak-tie-but-useful content in separate, clearly labelled sections, so interesting content from a stranger never looks like something to act on. Show the relationship numbers per entry so the ranking is auditable and the user can correct it.
Summarize Chinese posts in Chinese and English posts in English. Say what a post is; never repost it verbatim.