Research Xiaohongshu notes and distill them into skills
An agent drives Chrome to search keywords, open top-liked notes, read screenshots and distill findings into reusable skills in a knowledge base. Task: The author developed rednote-bootstrap, an open-source skill pack for AI agents that automates Xiaohongshu research. It uses agent-browser to control Chrome, splitting queries into sub-dimensions, searching keywords, and screenshotting top-liked notes. A vision model extracts text from images, which is then deduplicated and cross-checked to create reusable structured skills. The system saves these to a knowledge base for future zero-latency retrieval.; tools: custom-built or unnamed agent. Human role: Source evidence is unavailable regarding specific human roles or manual intervention steps in the provided text.; limitations: A full research run takes 5-10 minutes, constrained by 5 search rounds, 30 notes, and a 15-minute timeout. It requires initial user login via QR code, with state persisted in xhs-auth-state.json. The system relies on visual models to read image-based content, as Xiaohongshu lacks a public search API. Human review is suggested to verify distilled skills against original notes..
Tools
custom-built or unnamed agent
Published on Agent Work Atlas:
Overview
An agent drives Chrome to search keywords, open top-liked notes, read screenshots and distill findings into reusable skills in a knowledge base.
Workflow
The author developed rednote-bootstrap, an open-source skill pack for AI agents that automates Xiaohongshu research. It uses agent-browser to control Chrome, splitting queries into sub-dimensions, searching keywords, and screenshotting top-liked notes. A vision model extracts text from images, which is then deduplicated and cross-checked to create reusable structured skills. The system saves these to a knowledge base for future zero-latency retrieval.
Reported result
For a publishing process query, the agent analyzed 6 top notes (highest with 4479 likes) and generated a comprehensive guide covering 2026 rules and timing. A second skill on account warm-up was distilled from 4 notes. The author reports that the system successfully creates executable guides from real creator experiences without manual compilation, though specific long-term accuracy metrics are not provided.
Human role
Source evidence is unavailable regarding specific human roles or manual intervention steps in the provided text.
Limitations
A full research run takes 5-10 minutes, constrained by 5 search rounds, 30 notes, and a 15-minute timeout. It requires initial user login via QR code, with state persisted in xhs-auth-state.json. The system relies on visual models to read image-based content, as Xiaohongshu lacks a public search API. Human review is suggested to verify distilled skills against original notes.
Try it
Act as an AI agent using the rednote-bootstrap workflow to research [replaceable_topic] on Xiaohongshu. First, perform a dry run: search for [replaceable_keywords], open the top 3 notes, and screenshot them. Use a vision model to extract text. Then, deduplicate and cross-check findings to draft a structured skill. Before finalizing, ask me to confirm the accuracy of the extracted points against the screenshots. If confirmed, save the skill to [replaceable_output_location]. Ensure the output includes source links and confidence levels. If the vision model fails to read text, report the error and stop. Do not publish or spend credits without my explicit approval.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.