For most of the history of software, the keystroke was the unit of
work. Engineers typed code; the better engineer typed code that
compiled the first time. Tooling improved; languages got safer;
editors did more on your behalf — but the act, at its core, remained typing.
AI coding moved the unit of work. The instruction is now the unit;
the code is the diff the instruction produced. Cursor, Claude Code,
Aider, Continue, Cline, Windsurf, GitHub Copilot Chat, Codeium, Zed —
every modern AI coding surface accepts a natural-language prompt and
ships a code change in response. The clarity of the prompt is the
ceiling on the quality of the change.
Try a prompt with a goal, a file path, one constraint and an
acceptance check. After dictating, confirm that the draft preserves
each requirement. Compare the complete process with typing, including
corrections and the follow-up needed after the agent responds.
Halopen is the Mac voice layer engineered for that loop. Not a
dictation tool retrofitted for AI coding — a dictation tool whose
wedge is precisely what AI coding needs.
Read the long-form essay
for the cost-of-articulation thesis applied to Claude Code.