Examples now constrain it
An Anthropic engineer says the newest models steer better with a smaller prompt, so the team cut about 80% of Claude Code's system prompt and deleted the worked examples that used to guide it.
For years the received wisdom in prompting was: show the model what you want. Give it a few examples, spell out the rules, and tell it plainly what not to do. At a fireside chat now released as a video and annotated transcript, two people from Anthropic's Claude Code team described doing the opposite. For the current top-end models they trimmed roughly 80% of Claude Code's system prompt, replacing hard 'do not do this' rules with lighter contextual guidance and, most surprisingly, removing the in-context examples entirely.
The reason cuts against the old wisdom. The examples were not helping the newer models do the task; they were capping them. An engineer who ships the prompt said removing them helped because the model was more creative than the examples the team had written for it. Older models still get the fuller, more prescriptive prompt. The instructions a model needs, in other words, now depend on how good the model is, and the better it gets the less scaffolding it wants.
There is a tension the talk does not dwell on. This same round of changes to Claude's prompts and harnesses is what many users experienced earlier as Claude quietly getting worse. What the team frames as an optimization and what its customers felt as degradation are, at least in part, the same edits. And the 80% figure is Anthropic describing its own product, so read it as a direction of travel rather than a measured result. Still, the direction is the point: if steering a capable model means writing less, a large chunk of accumulated prompt-engineering craft is aimed at models that have moved on.
The lenses
The facts
How this connects
Tap a node to open it