Moving a 35KB agent prompt to self-hosted Ollama: what breaks
A builder's HN write-up on quitting Claude for local models: a 35KB prompt eats 14% of context and thrashes in 3 minutes — and the discipline that fixes it.
Copy markdown
The context trap: 35KB in, 14% gone
The 35KB preprompt swallows 14% of the window before any work starts, and on a local 65K-token cap the prompt plus session history overflow fast. Within three minutes the agent thrashed — repeated tool calls, re-reading files it had already read, rewriting finished work. The lesson: frontier models' giant context windows were quietly masking a bloated prompt.
The fix: one job per prompt, and raise the context
Splitting the mega-preprompt into single-objective units stopped the thrash; so did swapping prohibitions for positive directives ('only do Y' instead of 'don't do X'), cutting tool calls per step, and handing off state through disk logs that re-read only the slice needed. Critical gotcha: Ollama's default context length is tiny — set it explicitly or your model silently truncates the prompt you just wrote.
What a self-hosting exit actually costs
The rig: an AMD Ryzen AI MAX+ 395 with 128GB of memory, 96GB handed to inference, running 27B abliterated open-weight models under Ollama. You trade the convenience of a hosted frontier model for control and privacy — and a real hardware bill. A grounded reference point if this week's Claude Code weekly-limit cut has you eyeing the door.
Elsewhere: Ollama models now run inside ChatGPT Desktop
Ollama v0.34.0 (Sept 10) lets you drive local open models directly from ChatGPT Desktop, with faster structured output on Apple Silicon and OpenAI-compatible tool search — a gentler on-ramp to the same self-hosting path the migration notes take the hard way.