AI Pulse · Under the Radar

A developer has shared a detailed account of moving a 35kb system preprompt from Anthropic’s Claude Opus to a self-hosted Ollama setup. That might sound like a technical footnote, but it gets to a practical issue for any company building AI into real workflows: changing models is rarely as simple as swapping one API key for another.

The prompt itself is part of the operating system around your AI. It contains instructions, rules, context and accumulated fixes for how the model should behave. When you move providers or bring a model in-house, that work may need to be tested, adjusted and maintained again. Outputs can change, edge cases can reappear, and teams may need to rebuild the checks they had come to rely on. The writeup drew 139 points and 74 comments on Hacker News, suggesting this is a pain point many technical teams recognise, even if it is still one person’s experience. You can read the original discussion.

For an owner, the lesson is simple. Don’t assess an AI provider only on price per token or headline model quality. Ask what you would need to move: prompts, workflow logic, evaluation tests, data connections, staff knowledge and monitoring. Keep those pieces documented and separated where possible. This is the kind of thing we build into an AI command centre, so decisions and dependencies are visible before a vendor change becomes an expensive surprise.