A human-in-the-loop research workflow distributed as an MCP server and an OpenClaw Skill. It turns an open-ended research task into a staged, resumable process from topic selection to review.
OpenClaw Skill: clawhub.ai/hy0235/skills/research-report-skill
The workflow separates research into explicit checkpoints:
source scan → topic selection → thesis and outline → evidence plan → drafting → review
At each stage, the user can approve, revise or pause before the workflow continues. Project state is persisted so a report can resume after interruption.
- Generate research directions from selected information sources
- Build a thesis-led outline and evidence plan
- Draft section by section with human approval gates
- Save and resume project state
- Surface missing evidence instead of inventing data
- Produce a final review checklist for logic, sourcing and consistency
The MCP server exposes one report tool with staged commands:
| Command | Purpose |
|---|---|
topics |
Generate and compare research directions |
outline |
Create the argument structure and evidence plan |
write |
Draft the approved outline section by section |
review |
Check logic, evidence and consistency |
status |
Inspect saved progress |
resume |
Continue from a saved stage |
- Node.js 18+
- Model Context Protocol server over stdio
- Stateful workflow core with local project persistence
- Human approval gates and explicit no-fabrication guardrails
- Distribution through GitHub and ClawHub
npm install
npm startDetailed Chinese usage and implementation notes are available in the repository documentation:
Built as a solo AI-native project through iterative product design, prompt/workflow design, implementation, testing and deployment with coding agents.
This is an experimental research workflow, not an autonomous source of truth. Users remain responsible for verifying citations, data and conclusions before publication or investment use.