BuildMidData / MLAuto-graded

Trim a conversation to fit the context window

An explicit priority order has to be honoured exactly: what is never dropped, what goes first, when to refuse.A conversation grows forever; the context window does not. Before every model call you have to choose which messages to send so the prompt fits the budget — without dropping the system instructions and without discarding the recent turns the user actually cares about. The shipped trim_to_budget gets this backwards: it keeps the oldest messages, drops the newest, ignores the system message, and can still overflow.

30 min

python

Start lab

You're done when

The result always fits within the budget.
The system message is retained rather than trimmed away.
Recent messages are kept and the oldest are dropped, not the other way round.
The remaining messages stay in their original order.
When everything already fits, nothing is dropped.
Multiple system messages are all retained.
When even the system messages cannot fit, a budget error is raised rather than silently overflowing.

What it measures

Context engineeringDeciding what earns space in a finite context window, and what is safe to lose.

Prompt engineeringPreserving the instructions and the recent turns that determine whether the next response is any good.

Impossible-case handlingRaising clearly when the budget cannot be met, instead of returning something that will fail at the API.

Your environment

Terminal + editorA full shell in the browser

Claude CodePre-authenticated, no key setup

Isolated microVMYours alone, wiped afterwards

Start this lab

30 min

Start labYou'll sign in to Anthropos to start. The lab opens in a private cloud workspace with Claude Code ready to use.

AI Labs is part of Anthropos Workforce. Not on your plan yet? Book a demo