Your AI Copilot Has a 'Mental Workday', Too

Ever been deep in a coding session and everything is just... flowing? You're building, fixing, making progress.

Then an hour later you ask something simple and the response is off. Slow, over-engineered, or just wrong. You rephrase. Nothing. You start wondering if the AI is broken.

I kept hitting this wall. Took me a while to figure out the problem wasn't the AI. It was me.

I kept forgetting to let my AI start fresh.

#The context trap: my "tired" AI

I'm a big fan of the Copilot CLI (which I run using my copilot_yolo alias). Lately I've been running it with claude-sonnet-4.5.

My typical workflow starts with jumping into an interactive session:

Now I'm inside the Copilot CLI, prompt changed and ready to go. My workflow looks something like this (imagine this as one continuous terminal session):

The AI spits out the classes I asked for, probably across two code blocks.

Task 1 of my PBI done.

Now imagine I keep going. Same chat. I add unit tests for the repository, then a validation class for the Person model, then I refactor SavePeople to be async, then I add logging...

That's the trap. I've stacked up dozens of small tasks in one session. The context is now carrying my model, my repo, unit tests, async logic, a validation class, and logging config. All of it.

Then I'm ready for the next PBI task: building the API.

Still in the same terminal window, I type:

Seems like the natural next step. But the AI is now wading through everything we did before. It's trying to answer through the lens of file I/O, test frameworks, my logging setup, and the FluentValidation library I mentioned 20 prompts back.

The output comes back weird.

It's not broken. It's just dragging too much baggage. It tries to be thorough by remembering everything, and ends up pulling in stuff from 30 prompts ago that has nothing to do with what I asked.

#The solution: let your AI start over

Here's how I think about it: an AI chat context is like a workday.

You start fresh in the morning. You work through 'Task A' (the repository), load all that context, then move to 'Task B' (tests), then 'Task C' (validation). By the time you hit 'Task M' (the API), your brain is full. You're still carrying everything from the morning.

I was doing the same thing to the AI. Asking it to start 'Task M' while it was still holding 'Task A' through 'Task L'. That's like handing someone a brand new brief at 4:55 PM on a Friday. It's not going to go well.

The fix is one command. Run it inside your Copilot session:

Context cleared. Clean slate. Ready to go again.

(If you actually want to pick up where you left off, you can use flags like --resume when starting a new session. The point is to be intentional about it rather than just leaving old context sitting there.)

Now I give it a clean brief for the new task:

The output is focused. Just the API, nothing else.

No logging. No validation. No file I/O. Just what I asked for.

This isn't giving up on your AI or admitting the session went sideways. It's just good hygiene. Related tasks can still deserve separate contexts. When you're starting something new, give the AI the same clean start you'd want yourself.

A new task deserves a new context.