You can just do things

• Open a terminal and use it without fear
• Understand what a CLI agent is and how it differs from a chatbot
• Use an agent to research, structure, draft, and iterate on a policy memo
• Evaluate agent output critically (know where it helps and where it misleads)
• Deliver an oral brief on a memo you built with an agent
• Know where to go to keep learning and tinkering

• Foundations — terminals, agents, the mental model (10 min)
• The landscape — tools, permissions, setup (10 min)
• Live demo — building a policy memo from scratch (15 min)
• Your turn — hands-on memo exercise (40 min)
• Oral briefings — stand up and defend it (20 min)
• Where to go from here — tinkering, tools, next steps (10 min)

• The text interface underneath every computer
• Everything you do with a mouse can be done with text
• GUIs are slow and bounded. CLIs are powerful.
• You type what you want in plain language. The agent does the work.

$ gemini
> Create a folder called "memo" and draft a 1-page
  policy brief on algorithmic regulation in Canada

✓ Created memo/
✓ Wrote memo/brief.md (1,247 words)
  — Problem statement, 3 policy options, recommendation

• You wouldn’t hand an intern a task and accept the first thing back without reading
• You give direction, context, and judgment. The agent gives speed and breadth.
• The quality of the output depends on the quality of your direction.
• Good: “The problem statement doesn’t explain why this is urgent for Canada specifically”
• Bad: “Make this better”

Claude Code

Anthropic · $20/mo (Pro)

curl -fsSL https://claude.ai/install.sh | bash

Codex

OpenAI · $20/mo (Pro)

npm i -g @openai/codex

Gemini CLI

Google · Free tier available

npm install -g @google/gemini-cli

OpenCode

Open-source · Free (BYOK)

curl -fsSL https://opencode.ai/install | bash

Learn the pattern, not the product. The skill transfers across all of them.

• The agent proposes actions. You approve or deny each one.
• Write a file? Ask first. Run code? Ask first. Search the web? Ask first.
• You can grant broader permissions as trust builds

Gemini wants to run: python analyze.py --input survey_data.csv

Allow?  [y] Yes  [n] No  [a] Always allow

> y

You are responsible for what runs on your machine and what gets produced under your name.

• Be specific: “the problem statement doesn’t explain why this is urgent for Canada specifically”
• If the agent gives you a citation, ask it to verify or check it yourself
• Try: “What’s the strongest counterargument to my recommendation?”
• If you’re stuck, ask the agent to critique your draft
• Remember: you can just do things. Want a comparison table? Ask. Want a chart? Ask.

Agents can read and write files. Limit the blast radius by starting them inside a dedicated project folder.

# 1. Create a project folder
mkdir ~/my-project && cd ~/my-project

# 2. Start the agent FROM that folder
gemini

# 3. Tell it: "Only work within this folder"
#    The agent will ask before touching anything outside

• MCP (Model Context Protocol) — plug in new tools: databases, APIs, Slack, Google Docs
• Custom instructions — teach the agent your preferences, your style, your project context
• Skills and workflows — build reusable prompts for tasks you do repeatedly
• Community — GitHub, Twitter/X, Reddit — people share workflows constantly
• New features ship every week across all providers

• The best way to learn is to try things and see what happens
• There is no “correct” way to prompt — experiment, iterate, be curious
• When something doesn’t work, ask the agent why and try a different approach
• Build up your own toolkit over time
• The people who get good at it are the ones who play with it