A Small Experiment Born from Laziness
I started leaving instructions as phone voice memos while on the move, then pasted the dictated text straight into Claude Code. I compared three methods across 20 card-news copy edits each. There was only one metric: how many revision requests it took until the job was done.
Results
| Method | Avg. revision requests |
|---|---|
| Typing it myself | 3.1 |
| Dictated memo pasted as is | 3.8 |
| Dictated memo + one rule line | 1.4 |
Why Raw Pasting Was Worse
Speech comes with reversals mixed in. Say something like 「Shorten this sentence, oh wait, keep the length and just change the tone」 and the AI often carried out the first half of the instruction too. Filler words and repetition were also read as instructions.
Three Steps to Try It
1. Talk freely into your phone's voice memo app. Do not tidy anything up beforehand.
2. Copy the dictated text and paste it into the Claude Code window.
3. Always add the line below at the very top of what you pasted.
The text below is a raw voice memo. Ignore filler words and repetition, and if I change my mind, treat the last thing I said as the final instruction. Before doing anything, summarize what you understood in two lines.
The biggest payoff came from receiving the two-line summary first. Misunderstandings became visible before any work was done, so there was far less to undo.
Wrap-up
Speaking fast does not make an instruction accurate. The voice memo is raw material, and that one line at the top works as a sieve for it. This is a small-sample experiment, so please take the direction rather than the exact numbers.