For vibe coding

Your agent is only as good as the paragraph you couldn't be bothered to type

Short prompts get short-sighted results. Long prompts get good ones — and take four minutes of typing you don't want to spend. Speaking the same paragraph takes about forty seconds. Waver records it, writes it down cleanly, and hands you text you can paste straight into Claude Code, Cursor, or whatever you're driving today.

The arithmetic of a good prompt

Typing runs at 40–70 words per minute if you're quick. Speaking runs at 130–150 and takes no thought. A 250-word spec is a four-minute chore typed and a ninety-second aside spoken. That gap is the entire reason your prompts are shorter than they should be.

Typed
“fix the login bug”
Agent guesses. You re-prompt four times.
Spoken
250 words of real context
Agent has the constraints on attempt one.
Cost
~90 seconds
Versus four minutes of typing you skip anyway.

What you actually talk into it

Speak the spec, not a sentence fragment

“Add auth” gets you a guess. What the agent actually needs is which provider, which routes are protected, what happens on token expiry, where the session lives, and what the redirect does on failure — which is 200 words nobody wants to type at 11pm. Say it out loud in forty seconds instead. Waver writes it down, you paste it, and the first attempt lands close enough to iterate on.

Describe bugs the way you'd describe them to a colleague

Reproduction steps are miserable to type and trivial to say. Talk through it: what you clicked, what you expected, what actually happened, what you already ruled out, which commit it started on. That last part matters most: the knowledge of what is not the cause is the first thing dropped from a typed one-liner, and the thing that stops your agent re-checking three files you already cleared.

Dictate acceptance criteria before the agent starts

The cheapest way to stop a long agent run going sideways is to state, up front, what “done” means. Speak the list — the empty state, the loading state, the error copy, what must not regress, which files are off-limits. Ask for a rewrite into a checklist and you have criteria you can paste at the start and re-paste at the end to check the work.

Capture the idea before the context evaporates

The fix for the flaky test arrives in the shower, on a walk, halfway through cooking. Record it on your phone in the ten seconds you still have it, and it's transcribed and waiting on your desktop when you sit down. Nothing gets lost to “I'll remember it”.

Turn a pairing call into a prompt

Meeting recording separates speakers, so a design discussion or a pairing session becomes a written record of who decided what. Summarise it into decisions and follow-ups, then hand the relevant part to your agent as context instead of reconstructing the argument from memory.

Sketches count as context too

Photo scanning pulls text and handwriting off a page, so the whiteboard box-and-arrow diagram or the schema you scrawled in a notebook can go into the prompt alongside your spoken explanation — rather than being re-typed, badly, as prose.

Four steps, none of them typing

01

Talk

Hit record and explain the change the way you'd explain it to a new teammate. Ramble. Backtrack. No length limit.

02

Get it in writing

Waver transcribes it — in English or any of 120+ languages, in that language's own script if you think out loud in it.

03

Shape it

Keep the raw transcript, or have it rewritten into a structured brief: context, goal, constraints, acceptance criteria.

04

Paste and run

Copy it into Claude Code, Cursor, or whatever you're driving. Keep the note — it's the spec you'll re-paste when the run drifts.

On the machine you code on

Waver runs on macOS and Windows, plus iPhone and iPad for the ideas that arrive away from the desk. Notes sync, so the spec you dictated on a walk is on the machine with the repo checked out.

macOSWindowsiPhoneiPad

Stop under-briefing your agent

Free tier, no card. Talk one spec through and see what a properly briefed run looks like.

Try Waver free →

FAQs

Does Waver plug into Claude Code or Cursor directly?

No — and it doesn't need to. Waver records and writes up your prompt, you copy it, you paste it into whatever agent you're using. That keeps it agent-agnostic: the same workflow works for Claude Code, Cursor, Codex, Copilot, Aider, or a chat window.

Will it mangle technical terms and file names?

Transcription handles ordinary code vocabulary well, but no speech model is perfect on identifiers. The realistic workflow is speak, skim, fix the two or three tokens that came out wrong — still far faster than typing 300 words. For exact strings like a stack trace, paste those in rather than reading them aloud.

Can it clean up my rambling into a structured prompt?

Yes. The raw transcript is one output; AI rewriting turns the same recording into a structured brief — context, goal, constraints, acceptance criteria — without you re-typing it.

What about pairing sessions and design calls?

Meeting recording captures the call with speaker separation, so you can turn a 40-minute architecture argument into a written summary and hand the decisions straight to your agent.

I sketch architecture on paper. Any use for that?

Photo scanning reads handwriting and diagrams from a photo, so a whiteboard sketch or a notebook page can become text you feed into the prompt alongside your spoken explanation.

Where does it run?

iPhone, iPad, macOS and Windows. Record on your phone on a walk, the transcript is on your desktop when you sit down.

Does it cost anything?

There's a free tier with no card. Premium unlocks unlimited AI from $4.13/mo.

Talk it through. Paste it in. Ship it.

120+ languages · works on every desktop you code on · free tier with no card.

Get Waver free →

Stop typing what you could have just said.

Free forever to start. Unlimited transcription, no card, no ads.