Bells and Whistles

Integrations, advanced workflows, and tools worth exploring

Once you have a basic AI workflow running, these are the things that make it dramatically more useful. None of these are required — pick what fits your work.

★ = I actually use this regularly. Italic notes are honest opinions, not endorsements.

Direct integrations (MCP)

The Model Context Protocol (MCP) lets agentic tools connect directly to external services. These are not copy-paste workarounds — the tool reads and writes to these services natively.

Currently available in Claude Code

Service What it can do
Gmail Read inbox, search messages, draft and send emails
Google Calendar Read events, create events, invite attendees, find free time
Google Docs Read, create, search, and edit documents
Google Sheets Read and write spreadsheet data
Todoist Create/update/complete tasks, search, manage projects
Slack Read and send messages in channels and DMs
GitHub Create issues, PRs, review code, manage repos

Example workflow: “Go through my inbox. Anything about hockey, add it to my Google Calendar.” This is real — it reads the emails, extracts dates, and creates the events.

Outlook limitation: AI agents cannot currently interact with the Outlook desktop app — Microsoft gates this heavily. Gmail and Google Calendar integrations are much more mature. There are workarounds for Outlook read access, but they require manual setup.

Setting up MCP

MCP servers are configured in your Claude Code settings. The Claude Code docs on MCP walk through the setup. Most integrations take 5–10 minutes to configure.

Skills: reusable workflows

Skills are saved instructions that tell the AI how to do a specific task. Instead of re-explaining your workflow every session, you invoke a skill and it runs.

See the Applications & Skills page for installable skills from this bootcamp and detailed workflow examples.

Writing and voice

Voice files

A voice file describes how you write — tone, sentence length, habits, what you avoid. The AI reads it and matches your style. See the Session 1 slides for how to create one.

Project instruction files

Files like CLAUDE.md or AGENTS.md load automatically when you work in a project. They store:

  • Stable preferences and permission rules
  • File path conventions
  • What the tool should and shouldn’t do
  • Project-specific context

These compound over time — each session gets better as the instructions improve.

Markdown and knowledge management

AI tools produce Markdown files (.md). A good viewer makes them immediately readable.

Tool Cost Why
★ Typora $15 Clean live preview, opens .md files like a word processor
★ Obsidian Free Vault-based notes with links, tags, search — good for storing prompts and session notes
Zettlr Free Writing-oriented Markdown editor for longer academic work
MarkText Free Simple, lightweight Markdown editor

Meeting capture and dictation

Tool What it does
★ Granola Meeting capture — searchable notes from conversations. AI can search past meetings for context.
★ Wispr Flow Dictation — speak and it types. Good for drafting on the go or narrating session notes.

Research and literature tools

Tool What it does
Elicit Literature review, evidence synthesis, structured research reports. Cool, but I find my own workflow faster.
Scite Citation context — see whether later papers support or contradict a claim
ResearchRabbit Citation-network exploration and related paper discovery. Same — impressive demo, but I haven’t stuck with it.
Refine AI-powered academic writing and editing for research papers. Lots of people love this. ~$50/paper. I haven’t tried it yet.

Workflow patterns worth trying

Once the basic tools feel useful, the next level is better workflow structure, not better prompting:

  • Reusable prompts saved as files, not typed from memory
  • Project instruction files that accumulate over time
  • Session handoff notes so you pick up where you left off
  • Separate sessions for separate tasks — don’t pile everything into one thread
  • Plan → execute → review → clear context → continue for bigger tasks
  • Compare two AI outputs instead of trusting the first one

Explainers and examples from others

Official documentation