Xiaohongshu multi-account content matrix

2026.02 – present 5 accounts × 5 content styles In progress

Running 5 accounts used to be more than one person working full-time could keep up with. Now it's compressed to checking a Feishu (Lark) table once a day — the machine handles execution, the person just handles topic selection and taste.

The business problem

Content-driven lead generation needs a sustained, multi-account, differentiated publishing rhythm. Running 5 accounts manually means repeating "find material → rewrite → add images → publish → reply to comments" dozens of times a day. All the operator's time goes into mechanical actions, leaving no time for the work that's actually valuable — topic judgment and account strategy.

The approach: turning content ops into a pipeline

An automation pipeline covering the full content lifecycle:

  • Collection——pulling trending topics and material from multiple sources
  • Filtering + AI rewriting——rewritten separately for each of the 5 account personas to keep the styles distinct
  • Image generation——AI-generated images matched to the platform's tone
  • Publishing——multi-account isolation via BitBrowser, driven by Playwright/CDP
  • Comment replies——automated engagement to keep accounts active

The whole pipeline runs on Lark Base (Feishu's Airtable-style database) as the operations hub: content status, scheduling, and account health all live in one table. The operator controls the pipeline just by clicking statuses in Feishu — no code required. The stack is Vue3 + FastAPI + Playwright.

Collab table · ops hub Collect AI rewrite Image gen Scheduled post Comment reply 5 accounts × 5 content styles — pacing, frequency, and risk-control intervals all orchestrated from the table
Content ops is broken into five stages, hubbed around one table — the person only issues instructions in the table, never logging into each platform's backend.

The hard part: balancing automation against platform risk controls

  • Multiple accounts need to be fully isolated — fingerprint browser environments, operating rhythm, and publish timing all have to differ
  • Platform selectors and page structures change frequently, so the automation scripts need constant maintenance
  • Rate limiting and humanized behavior patterns to avoid tripping risk controls
  • Failure alerts——a failed publish has to be known immediately, not discovered three days later when an account has gone quiet
The boundary of content automation is clear: the machine handles execution, the person handles taste. AI rewriting and image generation solve the throughput problem, but the final call on topic direction, account persona, and content quality still needs a person. Fully-automated junk content hurts an account more than no content at all.

Results

Day-to-day operation of the 5 accounts went from "more than one full-time person can keep up with" to "check the Feishu table once a day." What this project proves: I can translate a complete operations workflow into an automation system that keeps running sustainably, and keep it alive within the constraints of platform rules.

Related retrospective: 《How I Built an AI-Assisted Xiaohongshu Content Matrix System》