Zurück zum Blog

How to Dictate a Coding Prompt With Clear Boundaries

Speak the goal and constraints, paste exact references, and review what the coding assistant may do before you submit.

12. Nov. 2026Whisper One TeamWhisper One Team

You are looking at a settings screen where a button jumps when an error appears. You can explain the problem faster than you can type it, but “fix the layout” leaves an AI coding assistant guessing. Say what you saw, what you want to change, and what should stay the same. Then use the keyboard for the details that need to match exactly.

This workflow gives you a draft to review in an editable prompt field. Whisper One enters text; the coding assistant is a separate tool. Check its permissions and your team's rules before including project information or letting it take actions.

Say what you observed in one short recording

Start with a description a teammate could confirm on screen. “When the profile form shows an error below the name field, the save button moves off screen on a narrow window.” If you only saw it on one device or in one test, say so. Mention the state that triggered it instead of guessing which component caused it.

Then add the outcome: “Keep the button reachable at a narrow width and keep the error visible.” This gives the assistant two things to check. “Make the form look better” gives it room to make design decisions you never meant to delegate.

If the behavior is intermittent or hard to repeat, first write down the exact steps. The guide to dictating a reproducible bug report can help you separate the observation from your theory about it.

Give the assistant a boundary it can follow

Say what belongs inside the task. In this example: “Inspect the profile form and its layout tests. Keep the existing validation rules and button labels. If the problem appears to require a shared component change, show me why before making it.” That last sentence gives the assistant a clear stopping point.

Use the permission level you actually want. If you are exploring an unfamiliar repository, start with “Read the relevant files and propose a small change. Do not edit yet.” If edits are allowed, say where. A request to verify a fix is not a request to deploy it. Make that distinction explicit whenever the tool has access to publishing or production commands.

Paste exact references after the narration

Add file paths, test names, error messages and any permitted screenshots by hand. Copy them from the source. If the file is unknown, say “find the profile form route” instead of inventing a path. Treat pasted logs as evidence for the task, and include only the smallest excerpt you need. Follow your organization's policy on sharing code, identifiers and customer data. A transcript can be sent to a speech service, and submitting it to a coding assistant shares it with a separate service. The dictation privacy checklist offers a useful review before you record.

Put the references after the instructions so you can review them separately. A short prompt might end with: “Verify in a narrow viewport with validation showing. Check that the error remains visible and the save button can be reached with the keyboard. Report which checks you ran.” If you know the exact test command, paste it in; otherwise ask the assistant to identify the appropriate test first.

Review the prompt before submitting

Read the final draft and underline the action verbs in your head: inspect, edit, test, publish. Is each one something you are authorizing? Next, check any spoken negatives. An omitted “don't” can change the task. Finally, confirm that the file names and expected behavior still agree with the example you described.

Start with one low-risk change. On an eligible Mac, download Whisper One, sign in and follow the quick start. Dictate the goal and the limits into an unsent prompt, add exact references, and submit the reviewed text yourself. The AI prompt guide shows how to place the cursor and start a short recording.