The AI CLI Landscape

The shift from "AI as chatbot" to "AI as terminal agent" happened between 2023 and 2025. GitHub Copilot for the IDE existed since 2021, but the jump to shell-native agents — tools that can read files, write code, execute commands, and iterate on failures — represented a qualitative leap. The three tools in this course represent the state of that landscape as it matures.

Each tool was built by a different company, for a slightly different primary use case, and with a different set of bets about what developers actually need from an AI in the terminal.

Each tool processes your prompts through a distinct pipeline. Understanding these pipelines tells you why each tool behaves the way it does — not just what buttons to press.

Here is what a basic shell interaction looks like in each tool's syntax — notice how differently they handle the same task:

# Claude Code — start an agentic session in your project
claude
# Then type: "Find all TODO comments, group them by file, and create GitHub issues for the critical ones"
# Claude reads files, groups results, asks for your GitHub token, creates issues.

# Gemini CLI — same task
gemini
# Then type: "Find all TODO comments, group them by file, and create a summary report"
# Gemini reads files (up to 1M tokens), writes report.

# GitHub Copilot CLI — suggest a shell command to find TODOs
gh copilot suggest "find all TODO comments in this repo and print them with file paths"
# Copilot suggests: grep -rn "TODO" . --include="*.py" --include="*.js"
# You press Enter to run it.

# Claude Code — start a session (PowerShell, Windows)
claude
# Prompt: "Find all TODO comments, group by file, create issues for critical ones"

# Gemini CLI — start a session
gemini
# Prompt: "Find all TODO comments and write a summary report"

# GitHub Copilot CLI — suggest a PowerShell command
gh copilot suggest "find all TODO comments in this repo with file paths"
# Copilot suggests: Get-ChildItem -Recurse -Include *.py,*.js |
#   Select-String -Pattern "TODO" | Select-Object Path, LineNumber, Line
# You confirm and run.

The canonical benchmark for coding agents is SWE-bench Verified. It measures whether an agent can autonomously resolve a real GitHub issue — no hints, no partial credit. It's the closest we have to a standardized measure of "can this thing actually fix my code?"

Architecture explains how each tool works. Philosophy explains why it makes the trade-offs it does. These three tools have meaningfully distinct world-views about what a "good" AI coding assistant should be.

All three tools support a "project instructions" file — a markdown file you drop in your repository that tells the AI about your codebase, coding standards, and preferred behavior. This is one of the most important patterns in AI-assisted development: the AI reads your docs before it reads your code.

The patterns are nearly identical between Claude Code and Gemini CLI — which means the skills you learn writing one transfer directly to the other. Here's a side-by-side of what a real context file looks like:

# Project: Payment Service

## What this does
Stripe integration for subscription billing.
Python 3.12, FastAPI, PostgreSQL.

## Key rules
- NEVER commit .env files
- All DB migrations must be reversible
- Tests required for every endpoint
- Use pydantic v2 models (not v1)

## Run locally
```
uvicorn main:app --reload
pytest tests/ -v
```

## Deployment
- Staging: push to `develop`
- Prod: PR to `main`, requires 2 reviews

# Project: Payment Service

## What this does
Stripe integration for subscription billing.
Python 3.12, FastAPI, PostgreSQL.

## Key rules
- NEVER commit .env files
- All DB migrations must be reversible
- Tests required for every endpoint
- Use pydantic v2 models (not v1)

## Run locally
```
uvicorn main:app --reload
pytest tests/ -v
```

## Deployment
- Staging: push to `develop`
- Prod: PR to `main`, requires 2 reviews

Knowing the tools is only useful if you know when to use them. Here's the decision logic distilled from real production teams who have adopted these tools at scale.