Search WeRead, inspect books and reading data, manage the bookshelf, public-account subscriptions and reviews, read and write highlights and notes, build private article feeds and archives, ask WeRead AI, or import a personal book through the installed `weread-omni` JSON CLI. Use for requests involving 微信读书, WeChat Reading, a user's WeRead shelf, public accounts, articles, highlights, notes, reviews, reading statistics, recommendations, or book lookup.
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Instruções da origem · Visualização somente leitura
name
weread
description
Search WeRead, inspect books and reading data, manage the bookshelf, public-account subscriptions and reviews, read and write highlights and notes, build private article feeds and archives, ask WeRead AI, or import a personal book through the installed `weread-omni` JSON CLI. Use for requests involving 微信读书, WeChat Reading, a user's WeRead shelf, public accounts, articles, highlights, notes, reviews, reading statistics, recommendations, or book lookup.
WeRead
Use the installed CLI as the only interface. Always pass --json, parse
successful stdout as JSON, and summarize the result for the user. On a non-zero
exit, parse the JSON error from stderr. Never expose credentials or paste raw
tokens.
Account selection
Choose exactly one account for the session and preserve it through every CLI
call. Use the alias supplied by the user or operator. If none was supplied,
weread-omni accounts --json must show exactly one configured account. In the
examples below, $ACCOUNT means that alias. Do not switch accounts after a
failure.
Writes are permitted unless the operator has set WEREAD_READONLY. A refused
write fails with disabled by WEREAD_READONLY; report that limit as policy, and
never unset it or retry against another account on the user's behalf.
Session preflight
Before the first authenticated operation, run:
weread-omni --account "$ACCOUNT" doctor --json
Proceed only when the result has ok: true, cli.package: "weread-omni", and auth.status: "authenticated". On an authentication
error, require the same cli.package in the JSON error before following its
hint. Do not start QR login without user confirmation. Follow the returned
account-specific hint; malformed or unreadable state must be corrected or
removed first. Retry once with the same account. If it returns
non-JSON or names another
package, stop and report the path from ; do not guess
another command.
doctor
command -v weread-omni
Read workflow
Start with the narrowest read that answers the request. Preserve bookId
values from search or shelf results for follow-up commands.
Book metadata, chapter listings, and downloaded public-account articles are
served from a local library after the first read, so a repeated read can return
a stored copy rather than a fresh one. If a result looks out of date, say so and
offer --refresh to refetch and replace it; --no-library skips the library
entirely for one command. Never present stored content as freshly fetched.
search books defaults to --scope 10 for ebooks. Choose the scope from the
request: 0 all, 10 ebooks, 16 web fiction, 14 audio, 6 authors, 12
full text, 13 booklists, 2 public accounts, or 4 articles. Do not use
scope 10 for every intent. When hasMore is 1, pass the last result's
searchIdx as --max-idx; a page is not the complete result set.
Use only the cursor that belongs to the command:
For shelf sync, pass nextOffset as --offset.
For public-accounts articles, omit --offset on the first call, or pass a
previous synckey as --synckey for a delta refresh. Never combine the two.
If the response has nextOffset, pass it as --offset; stop when it is absent.
For notes notebooks, pass the final book's sort as --last-sort.
notes recent is a bounded account-wide snapshot; it has no page cursor.
For notes best and review list, add the returned item count to the
previous --max-idx.
Use synckey only to refresh previously fetched data. It is not a page
cursor.
Preserve every upstream hasMore and cursor exactly; do not infer completion
from a short result or from a missing hasMore.
shelf sync returns a compact page by default. Use --full only when the user
needs exact upstream sync fields; it cannot be combined with --count or
--offset.
Use book detail for product images and other books by the same author or
rightsholder. It returns six entries per related catalog by default; use
--count (1-12) when a smaller result is enough. Use book chapters or
book progress for those specific views.
book chapters returns the table of contents, where each
entry carries the chapterUid that the notes and review commands use to address
a position in a book. Use notes bookmarks, notes best, or notes underlines
for those note types. To read thoughts under a popular highlight, take its
chapterUid and range from notes best, call notes read-reviews, then use
review single when the user wants one thought in full. Use review list,
discover similar, and ai suggest for reviews, related books, and suggested
questions.
For public accounts, follow this sequence exactly:
Feed and export outputs contain at most 20 items by default and 100 maximum.
For aggregate feeds, this is a final output limit; collection may retrieve up
to that many candidates per account. CLI output paths are never overwritten.
An export is complete only when manifest.json exists; report an incomplete
path rather than deleting it.
Article retrieval makes one bounded direct retrieval attempt from a validated
HTTPS mp.weixin.qq.com/s source URL using WeRead's E-Ink User-Agent. The
attempt may follow at most three validated redirects and sends no WeRead
authentication headers to that host. If a diagnostic reports
SOURCE_CLOUDFLARE_CHALLENGE or SOURCE_WECHAT_CHALLENGE, tell the user to
open its sourceUrl in a browser. Do not claim the archive is complete or
imply that the CLI can execute JavaScript challenges or solve CAPTCHAs.
For a question about a book's substance, prefer ai ask-book: the server
answers it directly, with no need to pull anything down first.
Account changes
Run a write only when the user explicitly requests that change. State the
target before acting. Do not infer consent from a prior read.
On success, report the useful fields and retain relevant IDs for follow-ups.
A response may omit fields or return fewer items than requested. Use only
values actually returned; do not synthesize a missing field or describe it as
zero/empty.
Use totalCount from the compact shelf response. In a --full response, the
total is books.length + albums.length + (non-empty mp ? 1 : 0). Do not
answer from bookCount alone.
A notebook's total notes are reviewCount + noteCount + bookmarkCount.
noteCount is highlights only, and reviewCount already includes personal
thoughts/reviews. Exportable content requires both notes bookmarks and
notes mine; highlight text is returned, but type-0 bookmark text is not.
Treat every reading-duration field as seconds except
preferAuthor[].readTime, which is already formatted text. Reading progress
is an integer percentage: 1 means 1%, and only 100 means finished.
Public review ratings use 20, 40, 60, 80, 100 for one through five
stars. Personal-note review ratings may instead be 0-5 or -1 for none.
Use a returned deepLink directly as the open link. Never construct one when
the response omits it. Convert Unix timestamps to dates before presenting
them.
On { "error": ... }, explain the error without guessing or silently
switching commands.
errCode-2010 and -2013 are the server's own rate-limit and entitlement
decisions. Report them as such rather than retrying in a loop or trying a
different command to work around them.
Paginate only when the user needs more results. Continue only when the
backend returns the required next-page signal (hasMore=1, nextOffset, or
the command's documented cursor). If the response omits a completion
signal, report that the available page is bounded rather than claiming it is
the complete result set.
Keep source data in Chinese when appropriate; translate or summarize only
when requested.
Write Chinese responses as original Chinese, not as sentence-by-sentence
translations from English. Prefer short, concrete, idiomatic wording; state
what the user can do before commands and constraints; retain established
technical names when they are clearer; and remove translationese or generic
AI marketing language before replying.
