Thariq Shihipar traces the arc of Claude Code’s system prompt across model generations: “originally… the best practice for a system prompt was a small system prompt, few tools, and lots of examples.” As models got smarter, the prompt grew — “you can give it more information and more instructions, and they start following them… a larger system prompt with lots of examples and many tools.” But the arc reversed: “most recently, we found this new class of models want a smaller system prompt. The examples tend to constrain it because it’s actually more imaginative than the examples we give it.” The team “removed 80% of the system prompt for Claude Code” and shifted from “do not do this” instructions to giving context instead of constraints — “which was really necessary for the previous models” but has become actively counterproductive for the new one.
The mechanism matters more than the anecdote: examples don’t just teach, they also implicitly cap — they communicate “the space of acceptable outputs looks like these” and a sufficiently capable model reads that as a boundary rather than a floor. This is the same underlying claim as Scaffolding is tech debt against the next model — the bitter lesson applied to product building — code and prompt scaffolding built to compensate for a weaker model’s gaps becomes dead weight, or worse, an active constraint, once the model no longer has that gap — but it names the specific failure mode inside prompt design itself (examples-as-implicit-ceiling) rather than the broader “delete your CLAUDE.md” prescription. It also gives Build for the model six months from now, not the model of today a concrete diagnostic: if you catch yourself writing another “do not do X” clause, ask whether you’re guarding against a limitation the current model generation has already outgrown.