Migrating 35kb Claude Preprompts to Self-Hosted Ollama — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- The author migrated 35kb preprompts from Claude Opus and Codex to self-hosted Ollama on a 128GB AMD Ryzen AI MAX+ 395 (32GB reserved for host OS) after concluding frontier providers cannot verifiably protect session data
- Self-hosted Ollama exhausts context within roughly 3 minutes when running large preprompts: a 35kb prompt immediately consumes 14% of the system's 65k token context window, triggering repeated tool calls, file re-reads, and rewritten work
- The OpenAI Navier-Stokes incident is invoked as a precedent — when pressed, OpenAI's best legal answer on training-on-user-sessions was reportedly "cannot rule it out" — which the author treats as proof that self-hosting is the only auditable privacy option
- Frontier provider safety filters are criticized as the "Refusal Industrial Complex" for blocking legitimate security work by treating vulnerability discovery as "hacking," preventing defenders from proving exploitability
- The Single Objective Prompting (SOP) playbook recommends splitting preprompts into one-objective units, declaring agents declaratively in opencode (~/.config/opencode/agents), tuning Ollama's context length explicitly, and logging session state to disk for frequent handoffs
- Mean Tokens To Forget (MTTF) failure signals to log include identical back-to-back tool calls, multiple reads on the same file, agents restating objectives, tool-call parse failures, and high turn counts relative to file changes
- The author's thesis on context reframes large provider context windows not as a capability upgrade but as a custody trade-off: the same hardware that supports inference is what gives providers access to session data
Why it matters: The concrete finding for builders: a 35kb preprompt on a 128GB AMD system immediately burns 14% of a 65k context window, forcing any practitioner who wants IP-safe AI to redesign around Single Objective Prompting and disk-based session handoffs rather than the monolithic prompts that work on frontier APIs. The article's larger argument — that "large context windows" function as the vendor's data-access mechanism rather than a pure capability gain — gives technical cover to developers who treat their session transcripts as proprietary work product.
Ask SkimNews
