Three Proven Feishu CLI Customer Cases: Bug Cycle from 40 Minutes to 5
At the Feishu CLI Developer Day in Shenzhen (July 17, 2026), one slide kept coming back: Feishu CLI handles connection and execution; the LLM handles understanding and judgment. That pairing is exactly what turns AI from a Q&A toy into something that runs real business processes. This post walks through three production cases shared on stage — retail store inspection, automotive bug triage, and law-firm case management — and shows the actual commands and prompts you can copy to reproduce each pattern yourself.
The pattern underneath all three cases
Every case follows the same three-beat loop:
1. PULL → lark-cli reads raw data out of Feishu (messages, docs, minutes, Base)
2. THINK → your agent / LLM classifies, summarizes, decides
3. WRITE → lark-cli writes the result back (a report doc, a Base record, an IM message)
lark-cli is the hands. The model is the brain. Neither is useful alone for business execution — together they close the loop. Keep this loop in mind; the three cases below are just it wearing different clothes.
Case 1 — Retail store inspection (NaiErBao, Tailanis)
Two retail brands pull in-store service recordings every day through the CLI, auto-generate a service-inspection report, and push it back to the ops group. The result: store service quality improved on a continuous, normal basis instead of only when a manager happened to do a spot check.
The skeleton you can reuse:
# 1. Find yesterday's store-recording minutes
lark minutes +search --keyword "门店巡检" --as user
# 2. Pull transcript + AI summary for each
lark minutes +detail --minute-tokens <token> --summary --transcript --as user
# 3. Agent scores service quality, then writes a report doc
lark docs +create --content '<title>7月17日门店服务巡检报告</title>...'
# 4. Push the report link into the ops group
lark im +messages-send --chat-id <ops_group> --msg-type text --content "今日巡检报告已出：<link>"
The prompt that drives step 3:
"你是零售门店服务质检专家。下面是某门店今日的接待录音逐字稿。请按【迎宾及时性 / 需求挖掘 / 产品讲解 / 送客规范】四个维度各打 0-10 分，指出一个做得好的细节和一个必须改进的问题，最后用一句话给出改进建议。输出为结构化小节，不要客套话。"
Case 2 — Automotive R&D: bug cycle from 40 min to 5 min
An automotive company uses the CLI to auto-pull user feedback and doc content from Feishu, then lets the model classify bugs and organize requirements. The bug-fix cycle dropped from 40 minutes to 5 minutes.
The leverage here is that feedback already lives in Feishu — scattered across groups and docs. The CLI pulls it into one place, the model triages, and the result lands in a Base where engineers actually work:
# 1. Search all feedback messages mentioning a symptom
lark im +messages-search --keyword "刹车异响" --as user
# 2. Agent clusters & dedupes, writes each bug as a Base record
lark base record create --app-token <app> --table-id <tbl> \
--fields {"标题":"...","严重度":"P1","来源":"用户反馈群"}
The triage prompt:
"你是汽车研发的需求分诊工程师。下面是一批用户反馈原文。请：1) 把描述同一问题的反馈合并；2) 给每条问题标注【模块】【严重度 P0-P3】【是否安全相关】；3) 安全相关的问题单独置顶并标红。只输出可分诊的结构化结果，不要复述原文。"
The 35 minutes saved per bug is almost entirely the "find + read + dedupe + file" work the model now does. The engineer only makes the final call.
Case 3 — Law firm case management
A Shanghai law firm combines the CLI with the AI whiteboard to auto-generate a case-tracking panorama, then archives the case info into a Base with one click — cutting the manual labor of organizing cases dramatically.
The two write targets matter here. The whiteboard gives partners a visual panorama of every case; the Base gives the team a queryable system of record:
# Archive structured case fields into a Base
lark base record create --app-token <app> --table-id <tbl> \
--fields {"案件":"...","阶段":"举证","负责人":"...","下次节点":"7-30"}
# Generate / update a case-panorama whiteboard inside a doc
lark docs +whiteboard-update --doc <doc> --whiteboard <wb_token> --mermaid "..."
Steal this: the reusable recipe
Strip the industry away and every case is the same recipe:
- Pick one data source already inside Feishu (a group, a folder of minutes, a doc).
- Pull it on a schedule with
+search/+detail. - Write a domain-expert prompt that outputs a structure, not prose.
- Write the result back to a Base (system of record) and an IM group (visibility).
Start with the loop, not the tooling. If your report never lands anywhere a human acts on it, no amount of model cleverness will move the business.