Voice coding vibes: how to code by talking with Cursor, Claude Code and Copilot

“Vibe coding” means describing to an AI assistant what you want to build and letting it write the code while you guide, review, and correct the direction. Tools like Cursor, Claude Code, GitHub Copilot, or Windsurf have made this practical. The bottleneck is no longer writing code, but writing instructions. And voice is unbeatable here.

Why speaking beats typing for agents

A coding agent needs context: what the app does, what you tried, constraints, and what “done” looks like. Typed, such a message is three lines. Spoken, it’s twenty — and the difference shows in the agent’s first attempt.

Typed: “fix the login bug”
Spoken: “At login, if the user already has an expired cookie session, they get error 500 instead of being redirected to the login page. I think the problem is in the auth middleware, which tries to read the user before checking expiration. Don’t change the token format; mobile apps use it too. Add a test that reproduces the scenario.”

Workflow

1. Describe the task by speaking. Place the cursor in the agent’s chat and hold Ctrl+Win. Say what you want, why, what not to touch, and how you’ll verify the result.
2. Let the agent propose a plan. Read the plan and correct it by speaking: “step three isn’t needed, we already have that function in utils.”
3. Review the diffs. When you see something odd, say exactly what: “why did you move validation from controller to model? Put it back, the model uses it and imports from files where validation differs.”
4. Debug by chatting. Paste the error and explain what you noticed around it; the agent gets your hypotheses, not just the stack trace.

The dictionary is essential

Function, table, and library names are the nightmare of any dictation: “use effect”, “get server side props”, “postgres”. Add your project’s terms (useEffect, getServerSideProps, PostgreSQL, your class names) to Zicelo’s dictionary and they’ll be written exactly like that. Zicelo also keeps English terms intact within Romanian sentences, without “translating” them.

What not to dictate

• Actual code. Don’t dictate parentheses or semicolons; describe what you want, and the agent writes the code.
• Secrets. Don’t say passwords or API keys out loud — not in dictation nor the agent’s chat.

Command mode for PR and commit descriptions

When you finish, select the change list and say “write a short imperative commit message in English” or “turn this into a pull request description with What, Why, How I tested sections.” It takes five seconds.

Some well-working dictated prompts

• “Before you change anything, explain in five lines how the payment flow currently works, so I can check we understand the same.”
• “Make the smallest change that fixes the problem; don’t refactor anything around it.”
• “Write the failing test first, show it to me, then fix it.”
• “Find all places in the project that read directly from the orders table and tell me which would be affected if I add a required column.”

Try it for a day: say every agent instruction longer than one sentence. Download Zicelo for Windows.