Use a scratchpad boundary
Dictate into a plain text editor first. Speech recognition is optimized for language, while shells are optimized for exact symbols. Hyphens, slashes, quotes, backticks, dollar signs, pipes, redirection, wildcards, parentheses, drive letters, environment variables, and case can all change behavior.
Copy the reviewed line into the terminal only after checking the current directory and intended target. Keep Enter as a separate manual action. For multiline PowerShell, verify how line breaks, continuation, and quoting will be interpreted.
Increase review with the blast radius
Read-only inspection can tolerate easier iteration. File deletion, recursive moves, permissions, registry edits, package publication, DNS, cloud resources, deployments, and remote commands require exact target resolution and a rollback plan. Elevated windows deserve extra caution because dictation software and target process permissions may differ.
Prefer commands with dry-run, what-if, diff, status, or list modes. Avoid broad paths, unresolved variables, and unreviewed globs. Print the resolved target before a destructive operation and preserve recoverability where possible.
Keep secrets out of speech and transcripts
Do not dictate passwords, API keys, recovery codes, tokens, private keys, or connection strings. A nearby person, microphone buffer, clipboard, local transcript, screen recording, terminal history, or analytics layer can create an unintended copy.
Use the operating system’s secure credential flow or a password manager. Avoid passing secrets as command-line arguments where they may appear in process lists or logs.
Best uses for voice in a terminal workflow
Speak comments, README steps, issue reports, commit-message drafts, search intent, test descriptions, and a natural-language request to a coding agent. Let the agent or editor produce precise syntax, then inspect the resulting diff or command.
This division preserves the speed of speech while keeping syntax and execution in a deliberate review loop.
Non-negotiable
No “always listening” or voice activation mode should be configured to submit terminal text automatically.