Building a voice-first workflow that survives week one

Last updated 11 September 2026

Why it usually fails

Not accuracy. Three things, in this order:

1. People try to dictate everything at once. Including the things speaking is bad at — code, tables, short replies, anything with a character limit. Those fail, and the failures colour the whole experiment.
2. They dictate word by word, watching the screen, which is slower than typing and feels awful.
3. The same fifteen names come out wrong every day and nobody tells them this is fixable in fifteen minutes.

All three are addressable. The order below is the one that works.

Week one: three tasks only

Pick the three where speaking is unambiguously better and ignore everything else. Resist the urge to test the limits — you will find them, and it will teach you nothing useful.

That is it for two weeks. Everything else stays on the keyboard. The point is to build the reflex on tasks where it visibly works, not to prove dictation can do anything.

The four habits

1. Speak a whole thought before looking

The single most important one, and the hardest. Watching text appear pulls you into editing mid-sentence, which is the thing dictation is worst at. Speak the entire sentence or paragraph, release, then look.

It is also more accurate. Speech models use the surrounding words to decide what a word was, so a complete phrase beats the same words spoken one at a time.

2. Read it back. Every time.

Ten seconds. Not a sign that dictation failed — part of how it works. Dictation errors are grammatical and plausible, which means they survive a glance and a spell checker. The read-back is the only thing that catches "can" where you said "can't".

3. Teach it fifteen words

Keep a note for one week of every word you correct. It will be about fifteen proper nouns. Teaching them takes fifteen minutes and removes most of your corrections permanently. See building a dictation vocabulary.

4. Switch to the keyboard without guilt

"Voice-first" does not mean voice-only, and the people who get the most out of dictation are the ones who stop using it without thinking about it. Type the email address, the file path, the code, the three-word reply. Mixing is the skill, not a compromise.

Week three onwards: what to add

Once the reflex exists, extend it in this order. Each of these is a genuine gain; each is also slightly harder than the last.

1. Chat messages. Slack, Teams, Messages. Short and conversational. Watch for Enter sending.
2. Longer documents, as first drafts only. Speak fast and loose, structure by keyboard afterwards.
3. Comments and feedback. On pull requests, documents, designs. Where the explanation gets cut for time.
4. The thing you have been avoiding. The difficult reply, the overdue note. Dictation lowers the cost of starting, which is usually what the avoidance is about.
5. Snippets for the repeated blocks. Once you notice what you retype. See voice snippets.

What to keep typing, permanently

• Code, regular expressions, file paths, anything with symbols.
• Passwords — macOS blocks dictation into secure fields anyway, correctly.
• Email addresses and URLs. Faster typed, and much more reliable.
• Figures, reference numbers, dates and times. A misheard digit looks entirely plausible.
• Headlines, subject lines and anything with a character limit. Compression is not a spoken skill.
• Structure: tables, nested lists, formatting.
• Anything under four words.

A day that actually works

1. First thing Clear the replies that accumulated overnight. Four of them, ninety seconds each, before the day starts.
2. Between meetings Dictate the note from the one that just ended, while it is still in your head.
3. When working with AI Speak the prompt. Sixty words instead of nine, and read the difference in the answer.
4. Mid-afternoon The first draft of whatever needs writing. Loose, fast, unstructured. Fix it later by keyboard.
5. End of day The handover or the update. Two minutes of speaking rather than a paragraph you never wrote.