dongbeixiaohuo/writing-agent

What it solves

Writing Agent transforms the process of creating content from a single-prompt "one-shot" generation into a collaborative, iterative co-creation process. It prevents the common issues of AI-generated text feeling mechanical ("AI flavor"), lacking factual accuracy, or ignoring the author's specific voice and intent by introducing structured discussion, multi-stage review, and human-in-the-loop confirmation.

How it works

The system uses a multi-agent orchestration approach where different specialized roles collaborate to produce a high-quality piece of writing. A "Writing Director" manages the overall process, while other experts handle specific tasks:

• Planning: Research and topic planning experts organize materials and structure.
• Writing: A lead writer creates the first draft based on confirmed outlines.
• ** uma Review**: Editors and "Reader Simulators" check for structure, value, and potential drop-off points.
• Refinement: A dedicated "De-AI" expert removes mechanical phrasing to preserve the author's voice.
• Verification: A fact-checking expert uses optional external search tools (like Tavily or Parallel) to verify claims.

Users can interact via a Windows desktop app (which manages state, versioning, and recovery) or as a Claude Code Skill. The process is conversational; the agent discusses the theme and materials with the user before writing, and requires user confirmation at each major stage.

Who it’s for

It is designed for content creators, such as those writing for public accounts (WeChat Official Accounts), opinion pieces, personal essays, and experience sharing, who have raw materials or ideas and want to refine them into professional, human-sounding articles.

Highlights

• Multi-Expert Workflow: Moves beyond a single prompt to a pipeline of director, writer, editor, and humanizer.
• Human-in-the-Loop: Requires natural language confirmation at each stage, allowing users to steer the direction and supplement materials.
• Fact-Checking Integration: Optional integration with Tavily and Parallel for real-time web search and evidence verification.
• Broad Model Support: Built-in presets for 99 model configurations across various providers (DeepSeek, OpenAI, Anthropic, Gemini, etc.).
• Version Management: Local storage of drafts and versions, allowing users to compare changes and export to TXT or HTML.