Your AI Agent Intern
Unlock the hidden superpower of your AI coding agent: it has read every single page of documentation for every tool you use. Here is how to manage that.
2026-01-17
Loading content...
Unlock the hidden superpower of your AI coding agent: it has read every single page of documentation for every tool you use. Here is how to manage that.
Treat your AI coding agent like a junior who has read every manual on the internet but still needs you to pick the tool. Brief it in chat before you let it write code.
I call AI agents "interns" when I want to lower my expectations. They miss context, they overreach, they need supervision. Fair.
If you stop there, you miss the upside. A human intern is still Googling how to center a div and reading your dusty deploy wiki. The agent has already ingested the React docs, the Stripe API reference, and every Stack Overflow thread from 2019. It will not look anything up. It will just pick the wrong API from memory.
Picture an intern who memorized the internet through last year. They know the obscure optional parameter on that AWS SDK method you touch once a year. They know the window-function syntax. They know six ways to configure your bundler.
I have fifteen years in this job and I still look up slice vs splice every time. The agent does not.
This documentation understanding is the real advantage. The failure mode is different: they have all that knowledge and still have no idea which tool fits your repo.
You cannot throw a ticket over the wall and hope. I treat the chat window like a whiteboard, not a shell.
When I start a payment flow, I do not open with "implement checkout." I say something like:
"We need a new payment flow. Stripe Elements. Match our existing checkout modal."
Then I ask how they would approach it, given our stack. Because they have read the docs, I might get:
That's where I steer. "We don't use Payment Element. Custom card styling. Individual Elements only."
I'm pruning their decision tree before any file changes.
Once we agree on the plan, execution is fast. The problem is bounded ("individual Stripe Elements, our Modal component"), so the agent can spit out the right props without wandering through alternatives you already rejected.
They don't need the docs at that point. You already picked the path.
This is the same briefing I'd give a human, and the same constraints I use elsewhere:
The agent is not a magic code generator. It's the most well-read junior on your team who still needs you to say which library, which pattern, and which file. Chat until the plan is boring. Then let it type.