Quick Start Checklist

Follow one question through the Research Workbench, from a selected source to a checked note. Then try its understanding check.

You can do this without an account or installation. Identify what goes in, what you save, and what you must check. If you are taking the course, follow your instructor's brief and keep your independent first reading separate from any comparison output.

Choose one output you need next: a reading note, a library search, or a comparison between accounts. Start with tools you already have. Reading and saving a note manually may be enough.

For the course, use the assigned mode: your own approved AI tool, a supplied AI output, human peers, or individual reading. A shared course setup supports its exercises; it is not a requirement for every researcher.

Consult the reference that helps with your next decision:

• Cognitive Blueprint Prompts - Structured prompt templates
• AI Model Reference Guide - Model families, reasoning controls, and sampling defaults
• Failure Museum - AI limitations and mitigation strategies

These explain prompting, model choice and failure checks. They are references to return to, not three readings you must finish before trying the walkthrough.

A host is the interface where you work with the model and, optionally, its tools. Keep an approved interface you already have if it serves the task. You do not need an agent to compare a response with a passage.

If you need another interface, choose one:

• Cherry Studio - the course's starting interface for comparing models
• Claude Code - a terminal agent
• Antigravity CLI - Google's terminal agent
• OpenCode - an alternative terminal host
• Kimi Code or Kimi Plugin - work with Kimi directly or within Claude Code or Codex

Follow that host's access instructions. Use API Keys for an assigned or provider key, or Vertex AI for an existing Cloud project. If you are running a course, choose the route first in Model access for a class. Reading, peer work and inspecting a supplied response do not require model access.

Use the task-based additions to choose what comes next:

• Zotero for a source library
• Research Rabbit or OpenAlex for source discovery
• Obsidian for linked notes, or Zettlr for writing in Markdown
• MCP connections when your host needs library or folder access
• PDF conversion when the source file needs it

Keep access limited to the intended inputs and destination. Test one connection with non-sensitive material before adding another. Check the output against the source, not just whether the tool returned a response.

Learn by example:

• Systematic Reviews Syllabus - 4-week course
• Session 1 is the review done by hand; Sessions 2-4 show AI workflows in action

Go deeper:

• Agentic Workflows - Multi-agent orchestration
• MCP deep dives (Sequential Thinking, Lotus Wisdom, Vox MCP)