A civilization can be measured by the questions it asks of its most advanced tools. We are asking ours to stop writing the introduction.
This is progress of a kind. The machine can explain almost anything, but you would like it to rename a file. Before doing so, it has prepared a short account of its understanding of file renaming, its approach to file renaming, and the benefits you may experience once the file has acquired its new name.
The Caveman project offers a deliberately terse communication style for coding assistants. Its premise is compression of prose rather than degradation of code. The repository also acknowledges that output brevity does not automatically reduce other kinds of token use, and that adding instructions can itself create overhead. Its numerical savings claims are project measurements, not ours.1
A service announcement about the service announcement
An answer can spend its first paragraph warming up, its second promising to help, and its third explaining that the subject deserves care. By then the reader could have found the setting themselves.
In human organizations, this wrapping serves familiar purposes. It signals attention, manages expectations, and provides evidence that somebody is working. An assistant that constantly announces its intentions can reproduce the same effect. The screen fills with activity. Whether the useful state of the world has changed remains a separate inquiry.
There are tasks where a progress update is essential. A long operation should not leave the user wondering whether it has stopped. There are also tasks where announcing the plan takes longer than executing it. We are still learning which is which, a discovery process apparently requiring considerable narration.
Brevity is not a security model
But a short message can accompany an expensive sequence of actions. A concise answer can be wrong. An omitted qualification can save words while transferring risk to the reader. The cost of an interaction is not identical to the length of its visible response.
This matters especially when the user needs to decide whether to trust a result. “Done” is pleasantly compact until it conceals the fact that only half the requested work succeeded. The missing sentence may have been the most valuable one.
A short wrong answer merely gets the reader into trouble faster. Keep the uncertainty that changes a decision; remove the ceremony around announcing it.
The luxury of being understood quickly
Those questions are older than AI. They are difficult precisely because the answer depends on the audience and the stakes. A warning about irreversible deletion deserves more care than a confirmation that a heading changed. A technical explanation may require detail that a status update does not.
A system that can write a long explanation ought to be able to judge when one is required. Otherwise, verbosity is doing the work of judgment, with considerably more tokens.
The industry has become very good at generating words. Its next test is whether it can bear to leave some of them out.