AI Coding Assistants
Autocomplete on steroids. You still drive.
▶ the lesson · narrated sketch video · tap for sound
the intro
AI coding assistants like Copilot and Cursor draft code from your context: the open files, your comments, the project around them. They shine at boilerplate and exploration, and they still need your review — you are the engineer, they are the intern.
fun fact ✦
The best assistant prompt is a good comment: describe the code you wish existed.
the docs
the deep dive
the ideas behind the video — concepts, diagrams, a walkthrough and the traps to avoid
An AI assistant is a brilliant intern with amnesia: it knows nothing except the context you hand it this turn, and it optimizes for plausible, not verified. That reframes the skill. Your leverage is curating context — the right files open, the pattern named — and reviewing output like it owes you money.
plausible is not the same as correct
key terms
context window
Everything the model can see this turn: your prompt, open files, chat history. It is finite, and crowded windows dilute attention.
the whole chat is one windowprompt
The instruction plus code you feed it. Specific goals with named patterns beat vague wishes every time.
// sort newest firsthallucination
Confident output referencing things that do not exist — APIs, flags, imports. Compilers and tests are the lie detectors.
invented function namesagent mode
A mode where the assistant acts: edits files, runs commands, iterates on failures. Powerful, and deserving of review.
it edits files and runs testshow it works
- 1Set the stage
Open the files that matter and close the ones that do not. Context is the raw material — garbage in, plausible garbage out.
- 2Describe the goal, not keys
Say what the code should do, the constraints, and an existing pattern to imitate. You are writing a ticket, not a spell.
- 3Generate small chunks
Ask for one function, not the whole feature. Small outputs are fast to review and cheap to reject.
- 4Make the tests vote
Run the code and the test suite before accepting anything. The compiler catches hallucinated APIs better than hope does.
- 5Review like it is a PR
Read the diff line by line as if a stranger wrote it — because functionally, one did. You sign off, you own it.
the code, explained
the comment-as-prompt pattern
side by side
blind accept
- fast until the first bug
- you ship code you cannot explain
- security holes slip through
- your own skills quietly rot
review it
- reading a diff is still fast
- you learn from every change
- bugs die before the commit
- you stay the engineer
Assistants trained on last year know last year: they will happily import a deprecated API or invent a config option that sounds right. When generated code fails to compile, do not debug the fantasy — re-prompt with current docs or a file using the correct pattern.
quick answers
Will this take developer jobs?
It absorbed the typing, not the job. Requirements, system design, and review are the work now — developers using assistants outproduce the ones who do not.
Copilot, Cursor, or an agent CLI?
They converge every month, so the brand matters less than the skill: context curation and review transfer to all of them. Pick one and get fast.
Is generated code safe to commit?
With tests and a read-through, yes. Watch for secrets, license-suspect snippets, and code nobody can explain — git blames you, not the model.
quick hits about AI Coding Assistants
short answers · interviews, tips, trivia & hidden gems