Review this change before I keep it
The review pass for code an assistant just wrote you, aimed at the mistakes AI-written code actually makes.
Dictation for developers
Brief your coding agent like a colleague, not a search box.
The prompt that works first time carries the context: what you tried, what must not change, how you will know it is fixed. Nobody types all of that. Say it, and Dictate Safe types it into your terminal, your editor or your AI chat, symbols and all. On the default setup, none of it leaves your computer.
25 free dictations · one-time purchase · macOS and Windows
A working day
One day, out loud.
Six moments from a developer’s day where talking beats typing. Each one is something the app does today.
09:10Team chat
Stand-up, before the call starts
YesterdayIshippedthebillingexport.Todayit’sthewebhookretries.
Kestrel# general# backend# releasesDirect messagesPriyaSam
# backend
P
Priya
Morning! Stand-up at 9:15. Post your update here if you can’t make the call.
S
Sam
Yesterday: the rate limiter. Today: search indexing.
Yesterday I shipped the billing export. Today it’s the webhook retries.
Say it the way you would say it on the call. The ums are gone before the words land, and they land wherever your cursor is: team chat, email, a ticket.
10:25Terminal, with your coding agent
The brief, with the context in it
Keeptheidempotencycheckasitis,anddonotchangetheexistingtests.
Your coding agent · ~/dev/kestrel
> run the webhook tests
14 passed
> where do we send webhooks from?
src/webhooks/deliver.ts, called by the queue worker in src/jobs/send.ts.
>The webhook handler gives up after the first failed delivery. Add a retry with backoff: 5 attempts, doubling the wait each time. Keep the idempotency check as it is, and do not change the existing tests. Add one that fails today and passes after the change.
Typing a prompt is slow enough that you cut it down to the request and leave the context out, then spend the next few rounds explaining what you meant. Saying it costs almost nothing, so the context goes in the first time.
On a 2018 Intel MacBook Pro, with everything running on the computer, a 186-word dictation came back as text in 5.8 seconds.
What you would typeadd retries to the webhook handler6 words
What you saidThe whole brief, in the terminal46 words
11:40Code Mode, in the terminal
Say “open brace” and get {.
ifisadminopen braceshowdashboardsemicolonclose brace
Welcome back
~/dev/kestrel
Tips
Hold your hotkey to talk
Speak in full sentences
Session
Offline, on this Mac
>
Listening
Dictation is built for prose, so it types the word “semicolon” and leaves you to fix it. With Code Mode on, around 110 spoken names become the characters they name, and four casing commands turn “camel case get user name” into getUserName, with pascal, snake and constant case too. AI cleanup and live typing step aside while it is on, so nothing rewrites your code.
open paren(
triple equals===
arrow=>
optional chain?.
spread...
backtick`
On seven of our own recordings it got 53 of 57 symbols right with the default speech model, and never once left the word “paren” sitting in the code.
13:15AI chat
The review you ask for every day
review
Review this change before I keep it
Payment webhook retries
Add retries with exponential backoff to the payment webhook handler. Give up after five attempts.
Here is the updated handler. It retries 5xx responses, doubles the wait each time starting at 500 ms, and gives up after the fifth attempt. A 4xx fails straight away, since waiting will not fix it.
TS
webhook.tsCode · TypeScript
Want me to add a test for the retry path as well?
Reply to Claude…
webhook.tsTypeScriptCopy×
import { post, sleep } from "./http";import type { WebhookEvent } from "./types"; const MAX_TRIES = 5;const FIRST_DELAY_MS = 500; export class DeliveryError extends Error { constructor(public status: number) { super(`Delivery failed: ${status}`); }} export class RetriesExhausted extends Error { constructor(public eventId: string) { super(`Gave up on ${eventId}`); }} export async function deliver(event: WebhookEvent) { let delay = FIRST_DELAY_MS; for (let n = 1; n <= MAX_TRIES; n++) { const res = await post(event); if (res.ok) return; // a 4xx will not get better by waiting if (res.status < 500) { throw new DeliveryError(res.status); } if (n < MAX_TRIES) { await sleep(delay); delay *= 2; } } throw new RetriesExhausted(event.id);}
Pasted into the chat you were in
Or press ⌥ P from any app
Dictate Safe’s Prompts page, then Claude: Prompts on the floating icon, one word, one click, and the whole prompt is in the reply box.
Click Prompts on the floating icon, type a word, click the one you keep. It lands in the reply box, ready for your diff. Over a thousand ready-made prompts come with a license, coding and servers among them, and the ones you write yourself come back the same way.
- Click Prompts on the floating icon
- Type a word from the one you want
- Click it, and it is pasted into the chat you were in
⌥ P on a Mac or Ctrl + ` on Windows opens the same search from any app.
Three from the library, written for this job.
Find the bug from this error
One cause, the line that caused it, the smallest honest fix, and the check that would have caught it.
Write the commit message and the pull request
Both, from the diff: a subject line under 72 characters, and a description a reviewer can act on.
15:30Your desk, empty
A two-hour job, and a cup of chai
Keeping awake, 1h 50m left
Welcome back
~/dev/kestrel
Recent activity
Rate limiter for the public API
Split the settings page
> Migrate the test suite from Jest to Vitest across all 142 files, fix whatever breaks, and keep going until the suite passes.
Migrating test files…38 / 142(esc to interrupt)
>
? for shortcuts
Dictate Safe
Keep awake is onLeave the lid open. Closing it still puts the computer to sleep. A locked screen is fine, whatever is running keeps running behind it.
Dictate Safe
Keep awake ends in 5 minutesOpen the Dictate Safe icon in the menu bar to add more time.
Claude Code on a long job. One click on the cup keeps the computer awake for two hours. Five minutes before the end it tells you, and the menu bar adds another hour.
Hand the agent a long job, click the cup, and go. Keep awake holds the computer awake for as long as you choose, warns you five minutes before the time is up, and lets go at 20% battery rather than running the laptop flat.
- 2 hours
- To start with. Or 30 minutes, 1, 4, 8 or 24 hours, or until you turn it off.
- 5 minutes
- The warning before the time is up. The menu bar adds another hour.
- 20% battery
- Where it lets go, after a warning at 25%, so it never runs a laptop flat.
- Lid open
- Leave it open: closing the lid still puts the computer to sleep.
17:50Pull request, in the browser
A description a reviewer can act on
Addsaretrywithbackofftothewebhookhandler…
Open a pull request
main ← retry-backoff
Retry webhook deliveries with backoff
Adds a retry with backoff to the webhook handler: 5 attempts, doubling the wait each time. The idempotency check is unchanged. Tests: a new case that failed before this change and passes now.
Create pull request
Say what changed and why, the way you would explain it at a desk. A browser is one more app with a cursor in it.
Any app, the browser too
Under an NDA
The dictation tool you can use under an NDA.
When a company restricts cloud AI tools, cloud dictation runs into the same rule: your voice, and everything you say about the code, goes to a server you do not control. Dictate Safe runs the speech model on your own computer. No telemetry, no analytics, and on the default setup, nothing sent anywhere.
Local, the default
Nothing leaves your computer. The speech model runs on it, and the filler words are removed on it.
Hybrid, if you choose
Only the text goes to the cloud provider you picked, for AI cleanup. Your voice stays on your computer.
Cloud, if you choose
Your recording goes to Groq, with your own key. We never see the audio, the text or the key.
Offline Lock is one switch that stops the app making any connection at all. Model downloads, update checks, signing in and cloud providers are refused while it is on, and the app says so in plain words. The app always shows which of the three it is in.
Measured, not promised
Our own numbers, from a 2018 laptop.
53 of 57Code Mode symbols right on seven of our own recordings, none left behind as words.
1.2 to 1.7 secondsFrom speaking to text, for a dictation of up to about 70 words.
About 12x realtimeThe default model on recorded audio: almost six minutes in about 30 seconds.
1,000+Ready-made prompts in the library, with coding and servers among them.
Speeds measured on a 2018 Intel MacBook Pro, everything on the computer, AI cleanup off. If yours is an M-series Mac or a recent PC, it is doing less work than ours was.
Questions developers ask
Does it work in my terminal and my code editor?
Yes. Dictate Safe types into whatever has the text cursor: a terminal, a code editor, an AI chat or a browser tab. Nothing is installed inside those apps; you hold your key and talk.
Will it rewrite my code?
Not with Code Mode on. While it is on, AI cleanup and live typing step aside, so what you say is typed as you said it, symbols included. Code Mode is off by default: switch it on for code and off again for prose.
Can it learn our function names and jargon?
Yes. Add them to your vocabulary list, and names, libraries and in-house terms come out spelled the way your team spells them.
Does anything I say reach a server?
Not on the default setup: the speech model runs on your computer, and you do not need an account to dictate. If you want cloud speed you can add your own Groq or OpenAI key; the app shows which mode it is in at all times, and Offline Lock stops every connection with one switch.
Which languages does it understand?
English. The speech models are English-only, which makes them more accurate at English than multilingual models of the same size.
Does it need a fast computer?
No. It runs on macOS 12 or newer, on Intel and Apple Silicon, and on 64-bit Windows 10 or newer. Every speed figure on this page comes from a 2018 Intel MacBook Pro. AI cleanup on the computer is optional and needs 16 GB of memory.
What does it cost?
One payment for a lifetime license that covers 2 computers, with updates and fixes to the features sold included, while we offer the app. Try 25 dictations first, with no card and no account.
18:30Lid closed
Try the part that matters first.
Twenty five dictations, no card and no account. It is the same app you would buy, running the same way it will on day thirty.
One time · 2 devices · 30 day refund
More stories
A different day, the same app.
PromptingOne brief, three AIsSay the prompt you would never bother to type.
Give me 3 options each with a rough budget, and keep it friendly, not salesy.
Under an NDAThree roomsThe dictation tool you can use under an NDA.
Question 12. When did you first see the revised contract?
