Qbit Wechat Review

量子位公众号内容复盘

Terminology

• 用户指标来源: user-provided WeChat backend screenshots, screenshot OCR results, or user-provided metric tables. This is the only metric source during ordinary review runs. Credential-based metric retrieval (wxdown-service, proxy capture, getappmsgext) is offline; do not ask the user to refresh WeChat credentials or provide app_secret.
• ordinary review run: a standard article review or same-topic analysis session, as opposed to first-run setup or explicit structure changes.

First-Run Setup

Before the first review run, ensure the local article discovery/body backend is available. Trigger setup when either condition is true:

• ~/Documents/Lark/tools/wechat-article-exporter is missing.
• http://127.0.0.1:3000/api/public/v1/download is unavailable.
• article list auth check at http://127.0.0.1:3000/api/public/v1/authkey returns invalid when title-only or same-topic discovery is needed.

→ Read references/wechat-exporter-backend.md §First-Run Install and follow it.

After ready, tell the user: 文章列表和正文抓取准备好了，指标请提供后台截图。

Entry Points

Entry A — WeChat Links

When the user provides mp.weixin.qq.com/s/... links:

1. Fetch article body/JSON via local backend (see references/wechat-exporter-backend.md §Fetch Article Body).
2. Read metrics from 用户指标来源. Field mapping in references/metrics-and-schema.md §Screenshot/Table Field Mapping.
3. Extract 负责人/作者 from body byline (<name> 发自 ... → write <name>).
4. Upsert Feishu Base record, upload readable Markdown to 附件, write AI复盘.

Entry B — Title List

When the user provides titles without links:

1. Use local wechat-article-exporter account search + article list first (see references/wechat-exporter-backend.md §Fetch Account And Article List).
2. Search the likely media account (量子位 by default; 新智元/机器之心 when provided), then match title against returned appmsgex.title.
3. Accept only high-confidence matches. If ambiguous, report uncertainty and ask for confirmation.
4. If resolved URL is a WeChat URL → continue with Entry A.
5. If resolved URL is a repost/media page → use for body and attachment only; metrics still require 用户指标来源.
6. Use ordinary web search only as a fallback when local exporter cannot resolve the article.
7. Do not stop after writing only title/link/date/read count.

Entry C — Same-Topic Analysis

Use when the user provides one topic plus three links (量子位, 新智元, 机器之心), or when a 同题分析 table row is queued.

Expected chat input:

同题：<topic name>
量子位：https://mp.weixin.qq.com/s/...
新智元：https://mp.weixin.qq.com/s/...
机器之心：https://mp.weixin.qq.com/s/...

Workflow:

1. Fetch body for all three links via local backend. Read metrics from 用户指标来源.
2. Upsert 量子位 article into 文章复盘.
3. Upsert 新智元 and 机器之心 into 竞品文章池.
4. Upsert one 同题分析 row, link the three records, write seven 同题统计表 rows and four structured analysis fields.
5. If any link is missing → mark 需确认, report before writing analysis.

For competitor discovery when links are missing, use local wechat-article-exporter account search + article list first with 媒体名 + topic keywords (e.g. 新智元 Siri Gemini 库克). Use ordinary web search only as fallback. Treat 搜狗微信 as last fallback due to captcha risk.

Base queue workflow and 同题统计表 dimensions/fields → see references/metrics-and-schema.md §Same-Topic V1 Schema.

Structured analysis fields on the 同题分析 record:

• 结论摘要: who performed best, where 量子位 sits. Mention reading scale as context only; main judgment from seven dimensions + secondary metrics.
• 标题差异: 标题 dimension only. Focus on information density, emotional intensity, core entities, numbers, contrast, share hooks.
• 内容差异: synthesize 起笔/叙事/证据/扩展/风格. Use 点赞率 for recognition, 在看率 for endorsement, 转发率 for shareability. Include metric-combination reading (e.g. high reads + low likes).
• AI复盘: practical next-step advice for 量子位 only. No repetition of the other three fields.

Completion checklist:

1. 文章复盘 has 量子位 article with body, metrics, attachment, AI复盘.
2. 竞品文章池 has 新智元 and 机器之心 with body, metrics, attachments, AI复盘.
3. 同题统计表 has exactly seven linked rows (结论/标题/起笔/叙事/证据/扩展/风格), strongest handling marked 🟨 .
4. 同题分析 record links all seven stats rows through 统计表, has all four analysis fields filled.
5. Read back the record and stats rows to verify links and content before reporting done.

Core Workflow

For any entry point, the end-to-end path is:

1. Read user's data source (Excel/CSV, screenshot OCR, or metric notes) → extract title + raw metrics.
2. Fetch article body via local backend or resolved URL.
3. Read metrics from 用户指标来源.
4. Extract 负责人/作者 from body byline.
5. Upsert Feishu Base record. Use 文章标题 or 文章链接 as identifier.
6. Upload readable Markdown to 附件.
7. Let Feishu formulas calculate secondary metrics and 内容质量分; do not manually write formula fields.
8. Write AI复盘 per references/ai-review-rubric.md.

AI Review

→ Read references/ai-review-rubric.md before writing any AI复盘.

Key principle: data comparison gives the clue, the article content explains the cause, the final judgment lands on content choices. Cite specific body evidence; avoid generic claims.

Field And Metric Rules

→ Read references/metrics-and-schema.md before changing fields, formulas, or view order.

Key constraints during ordinary review runs: do not modify Base structure; do not recreate deprecated fields; do not infer 发布位置 from title alone.

Required Companion Skills

• lark-base: Feishu Base structure, fields, views, records, formulas.
• Spreadsheet tooling: parsing .xls, .xlsx, .csv.
• Local wechat-article-exporter: account search, article list, article body retrieval.
• Screenshot/OCR or user-provided tables: metrics.
• Web search: title-to-URL resolution and competitor discovery.

Source Selection And Article Matching

• Prefer user-provided WeChat URL as authoritative source.
• For title lists, resolve via local wechat-article-exporter account search + article list before fetching body; use web search only as fallback.
• Accept near matches for punctuation/truncation/headline variants.
• If multiple candidates are plausible, compare publication date, topic entities, and opening paragraphs.
• Record uncertainty instead of pretending the match is exact.
• WeChat covers are authoritative over website covers.