Use voice where language is forgiving
Speak the problem statement, constraints, acceptance criteria, comments, commit-message draft, issue description, or a long AI prompt. Those forms benefit from narrative speed and can tolerate a correction pass. Dictating exact syntax is possible, but the correction burden grows quickly around braces, operators, case, filenames, and indentation.
A practical pattern is voice to scratchpad, then review, then paste into the terminal or editor. That extra boundary is a feature: it prevents recognized speech from becoming an immediate command.
Never couple recognition directly to execution
A shell can interpret one mistaken symbol as redirection, a wildcard, command substitution, a changed path, or a destructive flag. Elevated terminals increase the stakes. Password prompts and secure fields should not be dictation targets.
Keep Enter manual. Read the rendered command character by character when it mutates files, permissions, credentials, deployments, or remote systems. Prefer dry-run and status commands. Resolve paths before recursive operations.
Safety boundary
Voice can draft a command. A person should verify and execute it. Do not configure a workflow where a recognition result automatically presses Enter.
Speak structured prompts in passes
First state the goal and relevant context. Then list constraints, files or systems in scope, behavior that must remain unchanged, and how success should be verified. Pause before including secrets, customer data, proprietary source, or personal information. A local dictation step does not change the privacy terms of the AI service that receives the final prompt.
Review names, versions, dates, negations, and instructions such as “do not deploy.” Speech errors tend to be plausible words, which makes them easy to skim past.
Build a mixed-input workflow
Voice, keyboard, mouse, switch access, and screen-reader commands can complement one another. Keep focus visible, provide a stop shortcut, preserve unsent text, and avoid modes that make it unclear whether the system is listening. Fatigue and speech variability can change throughout the day, so controls should not depend on one input method alone.
For a workplace accommodation, assess the person’s actual tasks and environment rather than assuming a general dictation benchmark establishes suitability.