StopDisTrain/sdt-skills

What it solves

SDT Skills provides a structured, modular approach to content creation for Xiaohongshu (Little Red Book), moving away from single, long prompts that are prone to hallucinations and difficult to debug. It ensures that the research, strategy, and drafting phases are explicit and separate, allowing users to repair specific failures (like a weak title) without restarting the entire content production pipeline.

How it works

The project organizes the content creation process into 15 independently callable "Skills" divided into four layers:

• Evidence Layer: Handles account modeling, public-content research, and benchmark filtering to find learnable samples.
• Production Layer: Manages topic generation, scoring, and the actual drafting of content, including hooks and covers.
• Quality Layer: Performs quality assurance (QA) on script flow, audience resonance, and the detection of AI-writing signals.
• Delivery Layer: Converts approved content into publish-ready packages, including WeChat-compatible HTML.

To ensure consistency, deterministic tasks (like metric normalization and title length validation) are handled by Python tools, while judgment-based tasks are left to the Agent and the human operator.

Who it’s for

Content creators, social media managers, and researchers focusing on Xiaohongshu who want a professional, traceable, and quality-controlled workflow for content research and production.

Highlights

• Modular Architecture: 15 distinct skills that can be called individually or as a 10-stage workflow.
• Explicit Repair Paths: Failures in specific stages (e.g., positioning) can be routed back to the specific relevant stage (e.g., drafting) rather than restarting.
• Deterministic Guardrails: Uses Python scripts for repeatable tasks like metric parsing and HTML rendering to prevent AI inconsistency.
• Read-Only Research: Built-in safety boundaries to ensure research tools cannot perform actions like liking or commenting on the platform.
• CI Validation: Automated repository validation to ensure metadata and tool self-tests pass on every push.