Kinds of AI Coding Assistants
There are many AI coding products, and their names, features and prices change month to month. Learning one product's buttons will date quickly. Learning what kind of help each one gives, and how much checking each kind needs, will last. For Kitty you will probably use more than one kind.
The rule
The more the assistant can do alone, the more you must check.
Most assistants fall into four broad kinds, from least to most independent:
- Inline completion. As you type in your editor, it suggests the next line or a short block. You accept or reject a few lines at a time, so every suggestion passes in front of your eyes.
- Chat assistant. You describe a problem or paste some code into a conversation, and it replies with an explanation or code. You decide what to copy into your project, so you move every change by hand.
- Coding agent. It works inside your project: it reads files, edits several of them, runs commands and tests, and reports back. Some run inside an editor, some in a terminal. It can change many files in one go, which is powerful and easy to under-check.
- App builder. A hosted service turns a description into a running web app, often with hosting included. It is the fastest route to something clickable, and the kind where you see the least of the code unless you go looking.
None of these is "the best". A good vibe coder picks the kind that fits the task, then matches the checking to the kind.
The shortcut
Before using any assistant, ask two questions: "What can it touch?" and "How will I see what it did?" A completion touches one line and shows it to you. An agent can touch your whole project, so you must look at the full list of changes afterwards.
Fix a typo in one message
Tiny and local. Inline completion, or just type it yourself.
The common mistake is giving the most independent kind of assistant the least review, because "it ran its own tests". Its tests are part of what you need to check.
Drills
Answer fast — trust the rule.
Which kind of assistant can edit several files in your project and run commands by itself?
Why does this course teach kinds of assistants rather than one named product?
An agent changed eight files to add "settle up" and says all its tests pass. What does the rule say?
You want a clickable mock-up of Kitty to show your flatmates tonight, and you will throw it away tomorrow. Which fits best?
Remember
- Rule: the more the assistant can do alone, the more you must check.
- Shortcut: ask "what can it touch?" and "how will I see what it did?" before you start.
- Trap: giving an agent less review because it ran its own tests.