Lesson 1 of 3 · Terminal AI & Socratic Ops
What is a Coding Agent?
AI assistants can write code in seconds. That's exactly why you need rules: if the AI does the thinking, your brain doesn't grow.
This lesson is locked. You can read ahead, but finish earlier stops to earn progress.
Listen
Watch
Read
AI agents and chatbots
- Chatbots (ChatGPT, Claude, Gemini) answer in a browser chat.
- Coding agents (OpenCode, Gemini CLI, Claude Code, Codex) run in your terminal and can read — and change — your project files.
Instruction files
Agents look for special files in your project before they start:
| File | Read by |
|---|---|
AGENTS.md | OpenCode, Codex, Cursor and most others |
GEMINI.md | Gemini CLI |
CLAUDE.md | Claude Code |
These files provide instructions where the chosen tool supports them; check its documented discovery and precedence rules.
The Socratic method
Socrates taught by asking questions instead of giving answers. A Socratic tutor asks "What do you expect this line to do?" rather than handing you the fix. It feels slower — but it's how real understanding forms.
Good vs bad prompts
- Bad: "Write my whole website."
- Good: "What does this error mean?" or "Which CSS property controls spacing between grid items?"
Agnostic on your machine
On your own computer you can use any AI tool you like. The learning approach is provider-agnostic, but file support and permissions vary by tool. An instruction file is not a guaranteed lock.
Instruction files guide behaviour; they are not a security lock
Tools differ in how they discover project instructions, which names they recognise and whether other rules take precedence. Check your tool's current documentation. A file called AGENTS.md is a common convention, not proof every agent reads it or obeys it perfectly. Mirrored GEMINI.md/CLAUDE.md files are compatibility aids where the chosen tool supports them.
Launching an agent in studio-hq establishes a useful working location, but does not necessarily stop access to other folders or the network. Use the tool's actual approval, filesystem and sandbox controls where available. If you cannot enforce a boundary, treat access as broader than the project and do not expose private material.
Rehearse a harmless tutor session
Choose your own supported tool from the optional setup quest and follow its current official installation/sign-in guidance. Some tools need paid API access, some offer limited free options, and support changes. No paid provider is required to continue academy lessons using the built-in tutor.
Confirm which folder is open, place the course's instruction file at the documented location and ask one harmless conceptual question: “What is the difference between saving and committing? Ask me a question to check my understanding.” Then ask for a whole file and observe whether it follows the tutor instructions. This tests behaviour in that session; it does not guarantee future compliance.
If it proposes an edit or command you did not request, pause rather than approve automatically. Inspect the filename, command effects and current diff. Read-only exploration can still reveal secrets in ordinary source files, so do not provide credentials or a home-directory workspace.
Keep assistance portable and understandable
OpenCode is one option; other documented terminal tools may fit your machine and access. Learn to describe the goal, supply non-sensitive evidence, ask for explanations and verify results. A tool name is not the learning outcome. Before the instruction-file activity, locate the correct file, explain its guidance and demonstrate the actual tool's response without pretending the text is an enforced sandbox.
Key terms
- Coding agent
- An AI tool that can read and edit your project.
- AGENTS.md
- An instruction file most agents read automatically.
- Socratic method
- Learning by being asked guiding questions.
- Prompt
- The message or instruction you give an AI.
Summary — copy into your notes
- Supported instruction files guide agents; verify the chosen tool's behaviour.
- AGENTS.md works for most tools; copy it to GEMINI.md and CLAUDE.md.
- Ask AI to explain, not to write it for you.