Moving a 35KB preprompt from Opus to self-hosted Ollama burned 14% of context before the first turn

patrickmccanna.net·medium signalXBlueskyLinkedInCopy link

Patrick McCanna's notes, published 2026-09-13/14 and now on the HN front page, document a migration onto a 128GB AMD Ryzen AI MAX+ 395 with 96GB for inference and a 65k context window, where the 35KB preprompt consumed 14% of context immediately and left almost nothing for session history. The failure mode was thrashing, not refusal: within about three minutes the model repeated tool calls, re-read files it had already read and rewrote finished work. His fixes are concrete enough to copy, splitting preprompts into single-objective units, using declarative opencode agents over shell scripts, setting context length explicitly since Ollama's defaults are too low, and logging session state to disk for handoffs.