CompCode Academy

Lesson 1 of 6 · Free to read

Pick how much control to hand the AI

Pick how much control to hand the AI

Audience: Final-year students and engineers in their first three years
Prerequisites: You can run a program from a terminal and read a short Python or JavaScript function.
Time: about 20 minutes
What you will keep: a one-page "which mode, when" card for your own work
This lesson does not cover: installing a specific tool, or company AI policies.

Why this matters

"Vibe coding" is the name people gave to a real way of working: you describe what you want in plain English, an AI writes the code, and you steer. The speed is real. So is the hangover: code you cannot explain, functions that do not exist, and bugs that only show up after you merge.

The difference between those two outcomes is rarely the model. It is how much control you hand over, and how much you check before you accept. This module trains that judgement. Interviewers and tech leads look for exactly this: not "can you use AI" but "do you know when to trust it".

Three modes, one rule

Most AI coding tools offer the same three modes under different names. Cursor, GitHub Copilot, Claude Code and others all have some version of them, and the names and menus change often, so learn the modes rather than the buttons.

ModeWhat the AI doesGood forThe trap
Completion (inline, "tab")Suggests the next few lines where your cursor isBoilerplate inside a function you already understandAccepting suggestions you did not read
ChatAnswers questions and drafts code you paste or applyExplaining errors, comparing options, drafting one functionAsking with no file context, then trusting the answer
AgentPlans, edits several files, runs commands, retriesA clearly scoped change across a few files"Improve the code" with no boundary

The rule that covers all three: the smaller and more reversible the change, the more autonomy you can safely give. A one-line completion is cheap to undo. An agent that touched twelve files and ran a migration is not.

Choosing, with examples

You are renaming a variable inside one function. Completion, or just do it yourself. Any AI step costs more attention than it saves.

A test fails with a stack trace you do not understand. Chat. Paste the trace and the function, and ask for the three most likely causes before asking for a fix. You want to understand, not just to get green.

You need a new endpoint, its validation and three tests. Agent, but with a brief: which files it may touch, what "done" means, and that it must run the tests and show you the output.

You are changing login, payments, data deletion or production config. Drop down a level. Use chat to explore and write the change yourself, or read every line the agent proposes. These are the places where a confident mistake costs money or trust.

When to keep the keyboard

Using AI less is sometimes the skilled choice:

  • You are learning a concept for the first time. Struggle a little first; ask the AI to check your attempt, not to replace it.
  • The change is faster to type than to describe.
  • You cannot tell whether the output is right. If you cannot review it, you cannot own it.

Exercise: your mode card

Write a short card you can keep next to your editor. Five lines are enough:

  1. One task type from your own work where you will use completion.
  2. One where you will use chat, and the first question you will ask.
  3. One where you will use an agent, and the boundary you will give it.
  4. One area where you will always review every line.
  5. One situation where you will not use AI at all, and why.

Keep it. Lesson 4 adds it to your prompt log, and the lab in lesson 5 tests it.

Key points

  • Learn the three modes, not one tool's buttons.
  • Autonomy should shrink as the cost of a mistake grows.
  • If you cannot review it, do not accept it.

New free lessons and course updates. Unsubscribe from any email.