"Most skills have a short life" thesis, corroborated by Anthropic's own 80% system-prompt cut

A thread from someone briefed by Anthropic’s Claude Code team makes a striking claim: most of what we think we need to teach AI tools has a short shelf life. The argument goes that general workflows and standard practices get baked into the models themselves over time. What actually matters long term is the stuff only you know, your preferences, your company’s specific knowledge, the context no public training run will ever capture.

The kicker? Anthropic apparently proved this internally. They cut Claude Code’s system prompt by 80% because the model had already learned most of what the prompt was telling it to do. The original thread frames this as a shift in how we should think about configuring AI systems. If the model already knows how to format code or follow common conventions, you’re wasting tokens repeating it.

This matters if you’re building anything custom. Early adopters spent time writing long instructions for models, teaching them company style guides or standard operating procedures. That work isn’t useless, but it decays faster than people expected. The models catch up. What doesn’t decay is private information. Your client history, your internal terminology, the edge cases your industry runs into that never make it into a training corpus.

The practical takeaway is to focus your setup effort on feeding models what they can’t learn elsewhere. That’s the kind of thing we build into an AI command centre, a system that pulls in your actual context rather than re-explaining what the model already knows. If Anthropic can strip 80% of their own prompts without losing capability, you probably can too. The question is whether what remains is genuinely yours or just instructions the next model update will make redundant.