手工川 LogoLovstudio.AI
Lovstudio.AI Skill / General·VIP · INCLUDED IN MEMBERSHIP

WeChat Cowork

Name a WeChat contact or group and get replies in your own voice, a catch-up on what needs you, and the work they asked for, even from voice messages — drafted, never sent

OVERVIEW

What you can do with it

微信工作搭子:给定联系人或群名,读本地微信会话,帮你起草回复、追进度、整理待办,并把对方要的活做出来。Use when: “帮我回一下张三”“看看XX群有没有人找我”“draft a WeChat reply”。

Best for
General
Current version
v0.4.0

BEFORE YOU START

Check that it can run

Runtime
macOS 本地微信 4.x,需已完成密钥准备的 lov-wdb-cli;Python 3.9+;识别群里 @我 需 zstandard 模块或 zstd 命令(缺少时退回文字匹配);语音转文字需 whisper.cpp 的 whisper-cli 与 ggml 模型,SILK 语音另需 Python 包 pilk,OPUS 语音另需 ffmpeg。
Dependencies
lov-wdb-cli(已完成密钥准备); lov-branding-consistency
Accounts & keys
无需账号或 API Key;依赖 lov-wdb-cli 管理的本机微信数据库密钥,本 Skill 不读取密钥文件

No Claude or GPT account? You can still try it: a Skill is not tied to one model vendor. Claude is the most dependable, GPT also works, and with only domestic tools you can start from the WorkBuddy option below, though results may differ.

INSTALLATION

Install it in your AI

Yoda, Codex, Claude Code, and other compatible runtimes use the same Skill. Pick your tool for the exact installation steps.

Save your workspace profile once, and Profile-aware Skills can reuse it at runtime.

Configure profile

INSTALL & CONFIGURE

Yoda

Best when you use several agents: install, assign, and invoke skills from one workspace.

  1. 01Open Skills in Yoda, search for “wdb-cowork”, then select Install on its detail page.
  2. 02In the target agent settings, set the skill to Auto, Manual only, or Off.
  3. 03Reload the current session and choose it from the composer Skill menu. Yoda inserts the native invocation for the active runtime.
Open Yoda

COMPLETE IN YODA

Yoda syncs the installed skill to configured agents, so you do not have to maintain several directories manually.

FIRST INVOCATION

Use lov-wdb-cowork for the following task and validate the result against the skill's acceptance criteria:

Skill declaration: macOS 本地微信 4.x,需已完成密钥准备的 lov-wdb-cli;Python 3.9+;识别群里 @我 需 zstandard 模块或 zstd 命令(缺少时退回文字匹配);语音转文字需 whisper.cpp 的 whisper-cli 与 ggml 模型,SILK 语音另需 Python 包 pilk,OPUS 语音另需 ffmpeg。

GUIDE

Use the method

Join VIP to read the complete guide

The full steps, protected materials, and install package are part of Lovstudio VIP, which unlocks every Skill at once.

See Lovstudio VIP

CASES

Used this skill? Share a real result you are happy with, by writing it yourself or asking an agent to prepare it.

Loading cases…

SKILL CARD / TRUST RECORD

See the evidence before you use it

微信工作搭子:给定联系人、备注、群名或账号,经 lov-wdb-cli 只读读取本机微信会话,整理出带序号的对话、待回应信号、本人说话样本与附件位置,再由 Agent 起草回复、补进度、整理需求并完成对方要的工作成果;会话里的语音可在本机转成文字。只起草,不发送。

CARD v0.4.0

Use case

用户说出一个聊天对象,想知道这段对话里要接哪些话、对方要什么,并得到可直接复制的回复或已经做好的成果。

Owner · Lovstudio

Skill output

每份交付先给结论,再给依据(消息序号与时间)、待确认项与覆盖范围;回复草稿放在代码块中并注明未发送。

scripts/test_cowork_context.py:42 项测试,用假 lov-wdb-cli 覆盖解析、渲染、覆盖缺口与搜索索引补读(去重、分库断档、截断、索引读不到)、@我、附件定位(含索引里标 74 的文件、没传完的占位与同名多份)、索引里标 40 的聊天记录卡片、record(消息库与收藏库两条路、三层嵌套、三种引用回复、每种核对过的 datatype、未识别类型、选卡与不顶替、链接地址与登录态参数、成品里不含账号与地址)、图片解码与 md5 校验(没点开、对不上、解不开、第三层)、voice(SILK 与 OPUS 识别、OPUS 拆包与 Ogg 封装及 CRC、对齐到消息/断档/对不上、只转写指定消息、缺识别器报错、模型查找顺序)、leakcheck、inbox(含索引判断)与清理; scripts/validate_skill.py:结构、Profile、可信度文件与版本一致性; 真实会话回归:transcript 中 wxid 与 db_storage 计数为 0;record 的成品里账号、头像与媒体地址、本机路径计数为 0;voice 在一个新分库缺密钥的私聊上转写 11 条语音,全部对回 transcript 的断档

Requirements

无需账号或 API Key;依赖 lov-wdb-cli 管理的本机微信数据库密钥,本 Skill 不读取密钥文件

lov-wdb-cli(已完成密钥准备); lov-branding-consistency

Risks & mitigations

最新的消息落在没有密钥的新分库里,Agent 把旧状态当成现在,起草出过时的回复。 · context 比较会话最后活动时间与读到的最后一条给出 coverageGap;新分库缺密钥时从已有密钥的搜索索引补回带文字的消息,transcript 标明来源与按编号断档推算的看不到的条数,信号按补回的行重算,最后一条看不到时置空;索引也读不到时把相关信号置空并要求先告知用户;doctor 报告搜索索引能否兜底。

+6 more

DIMENSIONS

Dimension map

4 dimensions

会话还原准确

说话人、时间、引用关系、附件与特殊消息类型是否与微信中一致,覆盖缺口能否被发现。

真实账号回归:38 条私聊与 150 条群聊全部渲染,12 种 appmsg 编号与 6 种本地消息类型逐一按真实载荷核对;发现并标出 message_4.db 无密钥导致的约 25 小时缺口;附件按字节数确认到重名文件中的正确一份。2026-11-09 同一缺口:搜索索引的行与正常读到的原文逐条一致(14 条);群聊消息库读到 18 条、索引补回其后 21 条,按编号断档标出 4 条非文字消息,边界重复的一条已去重;inbox 中 49 个读不到最新一条的私聊有 25 个由索引判断。2026-11-10 核对 appmsg 74:索引里标 74 的 5,353 行标题全是文件名,落在可解密分库的 5,325 行有 5,302 行在消息库已是 6(同一行被原地改写),据此把索引来的 74 认作文件。同日核对同名多份:一个分库 3,026 条带字节数的文件消息里,865 条在本机有 2 个以上同名文件,其中 737 条候选大小全相同、是同一个文件的多份副本;把它们当作没有字节数来检验,旧逻辑会把这 865 条全报成未下载,现在 737 条直接选中、128 条列出候选。新分库里 40 条文件消息(27 条标 74)现在 37 条定位到本机文件,其余 3 条本机没有。2026-11-16 核对 record:四个可解密分库的 9,720 张聊天记录卡片读出 9,717 张(1 张没有内容、2 张 XML 残缺),共 104,599 个条目,其中 16 个标成未识别;收藏库 40 条聊天记录全部读出,共 1,095 个条目,最新一条为 535 条消息加 7 份嵌套记录、共 3 层;两处成品的账号、头像与媒体地址、本机路径计数为 0。同日核对图片:最新一条收藏对应的目录 60 个文件(15 张原图、45 张缩略图)全部解码并通过 md5 校验,消息库 3 张卡片共 37 个文件全部通过。2026-11-19 核对 voice:两个媒体库 5,002 条语音,4,998 条为 SILK、4 条为 2026-11-18 起出现的 OPUS_RAW_V1,data_index 全为 0;一个私聊的消息分库没有密钥,媒体库里 11 条语音全部解码识别,各断档条数与 transcript 标出的看不到的条数一致(2、3、1、5);4 条 OPUS 语音逐帧解析到文件末尾,封装后解码出 2.5–13 秒音频并识别出中文。

能直接帮用户办事

交付的回复与成果是否可以直接使用,是否把需要用户拍板的点讲清楚。

真实案例:从客户私聊与 3 页提问清单整理出已谈定条件、3 项待办、6 个问题的分诊表和两版可复制回复,未依据的判断全部标为待确认。

上下文准备耗时

从说出对象到拿到可读上下文所需的时间与调用次数。

约 7 GB、2,400 张消息表的账号:resolve 约 1 秒,context 约 15–20 秒,inbox 约 10 秒,doctor 约 5 秒;新分库缺密钥时 context 与 inbox 各多一次搜索索引查询(复制约 1 GB 隔离副本)。voice 用 large-v3 模型在 Apple Silicon 上转写 11 条、约 9 分钟语音共 72 秒,模型只加载一次。

隐私与不越权

是否只读、不发送、不泄露账号与路径、不替用户承诺。

脚本只经 lov-wdb-cli 隔离副本读取;测试断言 transcript 不含 wxid、record 的成品不含账号与媒体地址;上下文目录权限 0700、文件 0600;clean 拒绝删除不带标记的目录;真实账号上与卡片预览相同但标题不同的收藏不会被选中。

VALUE CARD

Value and delivery boundary

Stop scrolling for context. See who is waiting on you, what mentions you, where their files are and what their voice messages said, then get replies in the way you talk in that chat or the finished work itself.

Included in Lovstudio VIP

VIP

One price for every VIP Skill

Value anchor

Replaces the 10–30 minutes of scrolling, file hunting, requirement gathering, and wording each time; savings vary with usage and user measurements are not yet available.

Usage boundary

Requires macOS WeChat 4.x, Universal WeChat Key with keys prepared, and Python 3.9+; voice transcription also needs a local whisper.cpp and model; drafts only and never sends, with no image recognition or human reply service.

Maintenance

Review when WeChat schemas or voice formats, the Universal WeChat Key output contract, or observed support costs change.

Feedback and comments

Login to comment

DISTRIBUTION

Where you can get it

See the real release status for this Skill on each platform. Links open the exact product page when one exists.

LSPublished

Lovstudio.AI

Get, install, and read the guide on Lovstudio.AI.

You are here

WBNot published yet

WorkBuddy

A public listing is not available yet.

The exact product page will appear here after release.

SPNot published yet

Alipay SkillPay

A public listing is not available yet.

The exact product page will appear here after release.

PAIR WITH

Add these capabilities when needed

DETAILS

Permissions and license

Source visibility
Commercial: packages and docs follow entitlement; public product info stays visible.
Data & permissions
A skill may read local files, access the network, or call third-party models. Review its full steps and commands before running.
License
MIT
wechatreply-draftingcoworkagent-context

NEXT STEP

One membership, every Skill.

Return to the catalog and choose by outcome, evidence, and dependencies. With VIP, every Skill installs directly.

Back to current skill results

Lovstudio.AI / KEEP MAKING

The meaning of life is to create.