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

Patrick McCanna documented a migration onto a 128GB AMD Ryzen AI MAX+ 395 with 96GB for inference and a 65k context window, where the preprompt consumed 14% immediately and left almost nothing for session history (patrickmccanna.net). 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 copyable: split preprompts into single-objective units, prefer declarative opencode agents over shell scripts, set context length explicitly because Ollama's defaults are too low, log session state to disk for handoffs.