Give it the outcome, then let it work
A clear brief should remove questions rather than generate more of them.
Retrospective field note for 2026-07-06, reconstructed on 16 November 2026. The archive date is not a claim of publication at the time.
For a page: who is it for, what should they do, and what visual direction must survive? For code: which behaviour changes and how will it be checked?
That is enough to begin many tasks. The user should not have to specify every folder name and implementation step as well. Choosing a maintained component, following the repository and taking a straightforward route are part of the assistant’s job.
Ask when the answer changes money, access, data or client-facing behaviour. Do not ask merely because several ordinary implementations would work. Pick one, keep the change coherent and show the result.
If a requirement cannot be met, say which requirement. Quietly substituting a nearby task is worse than an honest partial result because it forces the person to discover the substitution themselves.
Sonnet 5 gives another model to test against that standard. A stronger assistant should take a compact brief further while preserving the intent. It should not require a longer brief just to avoid doing the wrong thing enthusiastically.
Anthropic introduced Claude Sonnet 5 on 30 June. Source.