Stream-of-Consciousness Prompting

Written prompts are expensive to produce and tend to fail in two ways at once: too short for the needed context and too specific about the solution. Speaking inverts that ratio. In ten minutes you effortlessly dictate several pages of context — history, failed attempts, preferences, exclusions — and naturally stay at the level of intent rather than implementation. That is precisely the point: modern models cope better with rambling, redundant context than with a terse, over-specified instruction that closes the solution space prematurely. Two conditions decide quality. First, transcription: without a project vocabulary, file, product and person names get mangled and the agent works on invented identifiers. Second, acceptance: whatever you did not specify precisely, you must check in the result. Spoken prompts shift effort from formulation to evaluation — they do not remove it.

For marketing and product teams this is the most obvious productivity improvement available, because it needs no new infrastructure. Briefings, campaign context, brand rules and history can be handed over by speaking at a quality nobody types voluntarily — and exactly that context decides whether results are usable.

Instead of typing a briefing in bullet points, a marketing lead speaks for fifteen minutes about audience, previous campaigns, unwanted phrasing and legal limits. The agent produces three variants from it; selection and correction stay with the human.

Dictation is old but long unsuitable for technical work, because technical terms and identifiers were reliably mistranscribed. With markedly better speech-to-text models and the option to supply project vocabularies, the approach became practical from 2025. By 2026 spoken prompting is common in developer circles for long mandates, while still impractical for short corrections.