Summary of Today’s Gen AI Classroom: Context in a File – CLAUDE.md and Coding Agents (5 Oct 2026)

Today split into two halves: configuring agent behaviour through files rather than code, and a hard look at what coding agents do to the way we work.

1. Project-Level Behaviour with CLAUDE.md

A markdown file at your project root that the SDK reads as standing instructions. Add “Always respond in Telugu” to it and every prompt in that project obeys, with no change to your application code.

Put project rules, architecture notes and goals here so the agent keeps its context across sessions.

2. Controlling Where Settings Come From

settings_sources decides which configuration files the SDK loads:

Level	Looks in
Project	the current project root
User	the .claude folder in your home directory

Pass an empty list to ignore both and get fully reproducible behaviour. That matters when you don’t want a developer’s personal settings leaking into shared output.

Keep project rules in CLAUDE.md at the project root, not in user settings. Project files travel with the repo; user settings don’t, and the same code then behaves differently on someone else’s machine.

3. Why the Claude SDK Feels Different

• OpenAI SDK: Python code calls the API directly.
• Claude Agent SDK: your code drives the Claude CLI, which talks to the model.

The CLI acts as a harness — the loop that runs the agent. With other frameworks you build that harness yourself. The CLI gives you a powerful one for free, at the cost of more tokens per task.

4. Claude Code

A command-line coding agent for writing new code, working through an existing codebase, or building something from scratch.

Setup

# macOS / Linux / WSL
curl -fsSL https://claude.ai/install.sh | bash

# Windows PowerShell
irm https://claude.ai/install.ps1 | iex

• You do not need Administrator mode on Windows.
• npm install -g @anthropic-ai/claude-code also works and needs Node.js 22+.
• claude not found? The install directory isn’t on your PATH. Add it rather than typing the full path each time, and remember Windows separates PATH entries with ;, Linux with : — that exact mix-up cost us time today.
• Run claude doctor to check your install and settings.
• Needs a Pro, Max, Team, Enterprise or Console account. The free plan doesn’t include it; Pro starts at $20/month.

Using it

• /init analyses your project and generates CLAUDE.md. On a new project it can scaffold a stack from a prompt, such as a gym-membership site in React.
• Plan mode: the agent reads the code, proposes a plan, you rework the plan, and only then does it write code. We demoed this on a Java Pet Clinic project, designing a community feature for pet owners.

An agent, by the same definition as always: an LLM for reasoning, tools to read code and Git history, and context from CLAUDE.md.

5. A Caution for Learners

Don’t lean on coding agents while you’re still learning to code. They’re productivity multipliers for people who already know what good output looks like. Without that, you can’t tell a correct plan from a confident wrong one.

Other tools in this space: GitHub Copilot, Codex, Google Antigravity, Amazon Kiro. None are required for this course, but knowing them gives context for the SDKs.

6. What Agents Are Changing

We ran a real demonstration: a UI test suite that normally takes a tester a full sprint was built in about 25 minutes using Playwright MCP and natural-language instructions, then wired into a CI/CD pipeline.

The same applies on the infrastructure side: Docker files, Kubernetes manifests and Helm charts now take minutes to generate.

Where this lands:

• The narrow role — only writing Selenium scripts, only assembling pipelines — is being absorbed into development work.
• Production and maintenance engineering is not going anywhere.
• As token prices fall under competition, adoption accelerates.
• The durable skill is judgement: knowing when the agent’s output is wrong, and being able to fix it.

If you have 7+ years of experience, 30 focused minutes a day on system design is the highest-return investment available right now.

7. Focus

Use the Pomodoro technique — 25 minutes of single-task work, then a break. Windows has a built-in focus timer. Procrastination is a scheduling problem more often than a motivation one.

✅ Action Items

1. Add a CLAUDE.md to your current project with its rules, architecture and goals.
2. Try settings_sources with an empty list and confirm the behaviour changes.
3. Install Claude Code, run /init on an existing project, and read the generated file.
4. Try Playwright MCP for UI testing.
5. No international card for Claude? Use Gemini’s free tier to replicate the same exercises.

⏭️ Next up: MCP, and deep agents for long-running tasks — whose design borrows heavily from Claude Code’s own patterns.