Process 2026-11-02
AgentC Blog

What would you tell Claude Code if you didn't have to type it?

The prompt you type and the explanation you would give a coworker are rarely the same length. A small test you can run on yourself shows what gets lost between them.

4 min read ~200ms if you're an agent Subscribe via RSS

Think of the last thing you asked Claude Code or Codex to do. Maybe you typed "fix the flaky checkout test" while the failed CI job was still on your screen, hit return, and went back to whatever you were in the middle of.

Now picture a coworker rolling their chair over and asking what you need. You wouldn't give them five words. What would you tell them?

An example: what we type

"fix the flaky checkout test"

What we'd say to a coworker

"The checkout test fails maybe one run in five, and only on CI. I think two runs are reusing the same order number, because it started when we added the retry. Leave the retry alone and just give the test its own order. Done means the CI job passes ten times in a row."

The second version is about fifty words, which takes roughly twenty seconds to say. It adds four things the five-word version left out: where the test fails, your best guess at why, what to leave alone, and how you'll know it's fixed.

You already knew all of it

None of that was news to you when you typed the short version. You left it out because typing it felt like a chore, and the agent has to guess at the rest. Sometimes the guess is fine. Sometimes the agent fixes a different problem, confidently, and you spend the next while undoing the detour.

We saw the same gap in a test of our own: Claude Code drafted a polite reminder that asked a sample customer for $200 they didn't owe, every time we asked in one quick line. The instructions you'd give a new bookkeeper caught most of them.

So why not just say it?

Talking to a computer has a history of small letdowns: waiting for the words to show up, words that come out wrong, and not knowing where the recording goes, which matters when you're describing a client's code or a customer's account.

So we tested Scribe Pro's speech engine on our own Mac. We gave it a recording about one sentence long, and the words came back right, with the speech engine done in a fraction of a second. The speech model ships inside the app and runs on the Mac's own chip, so the recording stays on your Mac.

Try the coworker test on your own Mac

Scribe Pro is the app we tried. Press the shortcut, talk, and your words land in whatever text field you're in, including the details you'd have skipped typing. Or say the job to Start Work, hand it to Claude Code or Codex, and watch every agent work in one window. Building on a Mac? Scribe Pro is $99.99 during early access, paid once.

Something to try today: before you send your next prompt, say out loud what you'd tell the coworker, and notice what's missing from the box.

Agents are patient listeners. Give them the version you'd give the coworker.

How we tested it

  • Setup: a release build of Scribe Pro on an M1 Max Mac, November 1, 2026, started with an empty home folder so nothing carried over from earlier use, using its bundled Parakeet v3 speech model.
  • What we timed: one 4.5-second recording (about one sentence) given to the app's transcription path, the same on-device engine live dictation uses. The engine reported 141 milliseconds from receiving the recording to finished text, and the text was correct.
  • What we did not time: live dictation from your voice, pasting the text into a window, or handing it to an agent. One recording, one Mac, one timed try; your Mac will give its own number.
  • The one wait: on that first launch, the speech model took about 36 seconds to get ready. A recording made in that window waits for it; ours was timed after.
  • Words a minute: people talk at roughly 130 to 160 words a minute in conversation, which is why 4.5 seconds is about one sentence and fifty words is about twenty seconds.