26:26Pat Simmons · Review
GPT-5.6 Soul vs Fable 5: I Gave Both Agents the Same Three Builds
Same prompts, three one-shot builds, two frontier coding agents — and one surprisingly clear winner.
July 11thThe argument in one line.
A venture-backed $2 billion dictation app can be cloned in an afternoon with Claude Code, using Apple's free on-device speech model instead of a paid API, showing the moat around consumer AI wrapper apps is thinner than it looks.
Read if. Skip if.
READ IF YOU ARE…
- You're comfortable directing Claude Code in natural language and want to see a full local macOS app built end to end from a single starting prompt.
- You're a Mac user paying monthly for a dictation tool like Wispr Flow and want a free, local alternative that never sends audio off your machine.
- You want to see how to compare two or three speech-to-text engines side by side before committing to one.
- You're curious how far a single Claude Code session can carry a real macOS app, from architecture plan to working GUI.
SKIP IF…
- You're on Windows — this build is Swift/macOS-only, though the creator says a separate Parakeet-based repo for Windows is coming.
- You want a production-ready polished app today rather than a rough one-session prototype.
The full version, fast.
Wispr Flow's push-to-talk dictation feature can be rebuilt for free using Claude Code and Apple's on-device speech transcriber, and this video walks through the whole build in real time. The creator starts with a single natural-language prompt describing the app he wants, gets an 8-piece architecture back (shell, HUD, hotkey, audio, speech-to-text, a backup engine, cleanup, and text injection), grants the macOS permissions the agent asks for, and fixes one Electron-specific insertion bug along the way. He then benchmarks Apple's native transcriber against NVIDIA's Parakeet model and against Wispr Flow itself, finds his clone is faster on both engines, and finishes by having Claude build a real GUI with transcription history and a custom dictionary.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:36
01 · Intro
Cold open needling Wispr Flow's $280M raise at a $2B valuation for a feature the creator says he just rebuilt himself; states the video's promise to build a local, free, faster clone using only natural language with Claude Code.
00:36 – 01:35
02 · The first prompt
Pastes one simple prompt into Claude Code inside Cursor: clone Wispr Flow as a push-to-talk dictation app, recommend architecture for his machine, start with a skeleton. Claude returns a plan.
01:35 – 03:16
03 · What the architecture actually means
Walks a slide-deck explainer of the 8-piece architecture Claude proposed: shell, HUD, hotkey, audio, Apple's native speech transcriber, Parakeet as a backup engine, a cleanup model, and text injection.
03:16 – 04:25
04 · Checking the build and granting permissions
Reviews Claude's build log and grants macOS Accessibility and Microphone permissions for the new app, renaming it 'Murmur YouTube' to avoid clashing with his existing personal clone.
04:25 – 05:08
05 · First test
Holds the hotkey and speaks into the app for the first time; the HUD renders and the transcript appears in the logs, but the text isn't landing at the cursor.
05:08 – 05:37
06 · The injection bug
Diagnoses why transcribed text won't insert: an accessibility API quirk reports success on Electron-based apps like Cursor even when nothing was actually typed. Claude fixes it and rebuilds.
05:37 – 06:50
07 · It works, and it's fast
Dictation now lands correctly and transcribes noticeably fast; the creator runs several rambling tests to stress-test speed and accuracy, calling the one-shot result pretty good.
06:50 – 08:14
08 · Downloading Parakeet from Hugging Face
Has Claude pull NVIDIA's Parakeet model directly from Hugging Face's public API via curl, skipping the Hugging Face MCP server so viewers without it can still follow along, as a second engine to benchmark against Apple's native transcriber.
08:14 – 09:58
09 · Parakeet vs Apple's speech model
Claude builds a small engine-comparison dashboard; head-to-head testing shows Parakeet edging out Apple's native transcriber on raw speed, 0.27s versus 0.32-0.58s, though both models mis-transcribe niche words like Claude Code.
09:58 – 12:07
10 · Head to head with Wispr Flow
Adds a third engine to the comparison by pulling live data from Wispr Flow's own database, then records the same rambling sentence through all three. Both homemade engines beat Wispr Flow on speed and are comparably accurate on the transcript text.
12:07 – 13:51
11 · Building a proper interface
Asks Claude to build a real macOS GUI, not just a menu-bar HUD, with a dictation history window and a custom dictionary, art-directed toward a 1980s tape recorder aesthetic: brushed aluminum, muted palette, no neon or vaporwave.
13:51 – 14:42
12 · The dictionary
Tests the finished Murmur YouTube app: transcription history now shows corrections, and adding Claude Code to the custom dictionary immediately fixes the model's recurring mis-transcription of that phrase.
14:42 – 15:29
13 · Wrap up
Confirms the clone works inside a real Google Doc, credits Wispr Flow's brand and marketing while reiterating the cost, ownership, and customization case for building your own, and points to the GitHub repo and blog post for anyone who wants to skip the build.
Lines worth screenshotting.
- A single natural-language prompt to Claude Code produced a working 8-component dictation app architecture: shell, HUD, hotkey, audio, speech-to-text, a backup engine, cleanup, and text injection.
- Apple's macOS 26 speech transcriber runs fully on-device, requires no API key, no local model download, and no per-minute cost.
- NVIDIA's Parakeet model, pulled from Hugging Face via plain curl, benchmarked faster than Apple's native transcriber in a head-to-head test, 0.27s versus 0.32s on the same sentence.
- In a three-way test against the real Wispr Flow, both the homemade Apple-based and Parakeet-based clones transcribed faster than the paid product.
- The one real bug in the build was an accessibility API quirk: some apps, including Electron-based editors like Cursor, silently report a successful text injection that never actually lands the text.
- Claude Code correctly built the macOS permissions flow, Accessibility and Microphone, without being told the specific system settings paths to use.
- Every speech-to-text model tested in this video struggled to correctly transcribe the phrase 'Claude Code,' consistently mishearing it as unrelated words.
- A custom pronunciation dictionary fixed the 'Claude Code' misrecognition instantly, correcting it in the live transcript the moment the word was added.
- Wispr Flow's own database read added roughly a second of artificial latency to its benchmark number, since the test had to ping Wispr Flow's servers instead of comparing local-only inference.
- Cloning a venture-backed app's core feature took roughly one and a half prompt turns to reach a fully working prototype.
A Paid App's Moat Was Thinner Than Expected
BUILD VS BUY
Cloning Wispr Flow's core dictation feature took one afternoon of natural-language prompts to Claude Code, and the free result matched or beat the paid original on both speed and accuracy.
02The first prompt
- A single plain-language prompt, description plus 'what would you recommend for architecture', was enough for Claude Code to scope a full macOS app rather than a rough script.
- Starting with 'give me a skeleton, then we'll build it into a proper app' let the creator validate the approach before investing in polish.
03What the architecture actually means
- Breaking a voice app into 8 named pieces, shell, HUD, hotkey, audio, STT, backup engine, cleanup, injection, turns an abstract clone request into a checklist anyone can follow, technical or not.
- Choosing Apple's built-in, on-device speech model over a paid API removes both the cost and the privacy tradeoff in one decision.
- Keeping a second, swappable speech engine in the plan from day one made a fair benchmark possible later without rearchitecting anything.
04Checking the build and granting permissions
- New macOS apps built by an agent still need the same manual Accessibility and Microphone approvals a human-coded app would require; Claude Code can't grant its own permissions.
- Reusing an app name that collides with an existing app is an easy trap; renaming early avoided confusing permission grants.
05First test
- The HUD and live transcription can appear to work perfectly while the actual output silently fails to land anywhere, so always verify the end result, not just the intermediate signal.
06The injection bug
- Some accessibility APIs report a successful text-insertion even when nothing was typed, particularly inside Electron-based apps, a silent failure that looks identical to success in the logs.
- Handing Claude the exact failure symptom let it diagnose and patch a platform-specific bug without needing that context volunteered up front.
07It works, and it's fast
- A working one-shot prototype, reached in roughly one and a half prompt turns, is a realistic bar for how far a single Claude Code session can carry a real macOS app.
- Stress-testing with fast, rambling speech during development surfaces edge cases that a single clean test sentence won't.
08Downloading Parakeet from Hugging Face
- A model can be pulled straight from Hugging Face's public API with plain curl commands, no MCP server or account setup required, when you explicitly tell Claude to skip the fancier integration.
- Keeping a fallback speech engine in reserve paid off the moment it was needed for a legitimate side-by-side comparison, rather than being a hypothetical nice-to-have.
09Parakeet vs Apple's speech model
- Raw benchmark numbers can be close enough that engine choice becomes a convenience decision, not a performance one.
- Every speech-to-text model tested consistently mis-transcribed a specific proper noun, a reminder that no off-the-shelf STT model is tuned for your specific vocabulary out of the box.
10Head to head with Wispr Flow
- A homemade, free clone built in a single session out-transcribed the subscription product it was copying on both speed and accuracy in this test, though the paid app's number was inflated by an unrelated database-latency call.
- Building the comparison tool to pull live numbers directly from the competitor's own database, instead of eyeballing it, is what made the faster-than-the-original claim demonstrable rather than just asserted.
11Building a proper interface
- A specific, referenced design direction gives an agent enough constraint to avoid the generic gradient-and-emoji look most AI-generated UIs default to.
- Asking a second, fresh Claude agent to draft the design-brief prompt before handing it to the coding agent is a way to get a sharper spec than writing it solo.
- Requesting design tokens first, before any views get built, keeps a UI redesign reviewable in one pass instead of scattered across dozens of component edits.
12The dictionary
- A user-editable pronunciation dictionary is the practical fix for a model's recurring blind spot on a specific word or phrase, cheaper than fine-tuning or swapping models.
- Corrections applying in the live transcript in real time, not just on the next recording, is what makes a dictionary feature actually usable day to day.
Terms worth knowing.
- STT
- Speech-to-text, the model that converts recorded audio into written words.
- HUD
- Heads-up display, the small floating on-screen widget that shows recording status while the app is listening.
- Push-to-talk dictation
- A voice-to-text pattern where the user holds a hotkey to record and releases it to have the transcribed text inserted at the cursor.
- Parakeet
- An NVIDIA-built local speech-recognition model, downloadable from Hugging Face, that runs entirely on-device without an API.
- Text injection
- The step where transcribed text is inserted into whatever app the cursor is currently focused on.
- AX API
- Apple's Accessibility API, which some apps use to receive programmatic text input; certain apps report success even when the text wasn't actually inserted.
Things they pointed at.
00:05productWispr Flow
00:10toolClaude Code
01:42toolCursor
02:20toolApple Speech Transcriber (macOS 26)
02:43toolParakeet (NVIDIA)
Lines you could clip.
00:05
“They just raised $280,000,000 at a $2,000,000,000 valuation for exactly what I'm doing right now using a tool I built in twenty minutes with Cloud Code.”
05:17
“That's the classic a x silent failure. Return success on electron apps.”
11:55
“Both of these models are officially faster than WhisperFlow, and our little clone is completely free.”
14:41
“If you're trying to save money, if you're trying to own your own data, if you're trying to customize this to your liking, you can see just how easy it is to do.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
analogystory
Whisper flow. Whisper flow. Whisper Whisper flow.
Whisper Whisper flow. Whisper flow is, I'm just gonna say it, completely unnecessary.
They just raised $280,000,000 at a $2,000,000,000 valuation for exactly what I'm doing right now using a tool I built in twenty minutes with Cloud Code. Look at that.
Perfect dictation. So in this video, I'm gonna show you exactly how to build your own version of Whisper Flow, how to run it completely on your own machine so nothing you say ever leaves your computer, how to make it look like a real app instead of a script in a terminal, and how to make it even faster than the one you're already paying $15 a month for.
This tutorial is designed for a complete beginner. All we're gonna be doing is talking to Claude Code, the natural language. By the end, you'll have your own version of Whisper Flow.
So not wasting any more time, let's dive right into the build, and we're gonna start with this prompt into Claude code. And for this prompt, we're actually gonna keep it really simple. All I'm gonna say is I want to clone Whisper Flow.
Look up what this is. I basically want it to be a push to talk dictation app. What would you recommend for architecture based on my machine?
We'll start with the skeleton, then build this out into a proper Mac OS application, and add some cool branding.
So let's just take that, copy it, and I'm gonna open up cursor here. I'm gonna open up Claude in here, and I'm just gonna paste this prompt. So you can see just how easy this is.
All I'm doing is I'm going I I don't know anything about the architecture or anything like that. All I'm doing is I'm saying, this is what I wanna build, look this up, come up with the architecture, and we'll just get to a working prototype really quickly, and then we can layer on things like building out an actual Mac OS application, updating some of the branding, know, the waveform that shows up, all that kind of stuff.
So Claude has a plan for us now, and it's laid out all these things, shell, h u d, hot key, audio, s t t, cleanup, injection. Now you really don't need to know what any of this stuff means, but for the sake of this being a proper tutorial, I just threw an explanation of these in this deck here. So this is what the architecture is.
So first, the shell, this is just the app itself. This is the gonna be the container that everything lives in, so we're building a proper swift application, how we're actually going to build the interface.
But this is just the heads up display, so when I actually just like you saw when I was transcribing initially, that little waveform that's gonna pop up, so we can see that it's actually being transcribed in real time. Hot key c g event tab, so this is just what button we're gonna hold to actually activate the transcription.
In this case, we're just gonna use f n. Audio is just actually the microphone that's gonna be picked up when we are talking. Then of course, we have STT, speech to text.
This is the model that turns your voice into words. In this case, we're just using Axe native speech transcriber. This is relatively new, I wanted to experiment with this.
It's built directly into Mac OS, so it's a 100% local, you're not calling an API, you're not paying for anything, and you're not even downloading a local model, just directly out of the box through Mac OS. And then just as an alternative, just in case speech transcriber is not good enough, we have Parakeet, which is Nvidia's local model that you can download really easily.
I'll show you how to do that through Hugging Face. We're gonna just have that as a backup in case speech transcriber is just not that good. And then Clean Up as well, this one's completely optional, but we're gonna have a small local model that's just gonna add punctuation, cut the ums, do that kind of stuff that WhisperFlow I know does really well.
Like I said, we absolutely do not need to do this. But just to see how good it can actually get to WhisperFlow, we're adding the second small model. And then finally, have injection.
So this is just making sure that the text is injected properly where our cursor is. So with that, let's check-in on the latest link architecture build. So it looks like it is done, I think.
Yeah. So you can see here it's just explaining Mac OS 26 ships speech analyzerspeech transcriber.
Apparently Parakeet v three beats it on English accuracy, so it did some research on that. We're gonna start with just speech transcriber just to see how good it is, and then we can test Parakeet later on, and then it has basically all the things that I already explained.
The HUD, hotkey, CG event tab, FN button, and then the only thing it looks like I need to do is just grant permissions. So system settings, privacy and security, accessibility, and then add what we're calling Murmur, this application that the agent just built. So we'll do that now.
I'll just go to privacy and security, accessibility, and then minor confusing thing, I already have a WhisperFlow clone that I've been using for a couple months now, also called Murmurs. We're just gonna change the name of this. We'll just call it Murmur YouTube.
Change name. Also, I just gave permissions. Do you have full access?
I just need to open the system settings. Open system settings. Okay.
Remember YouTube, we changed the name. Alright. Okay.
Should have full access. Now please test. Eight would like to access the microphone.
Allow. Try now. Alright.
So after allowing those permissions, it looks like this is working. Claude drove this with a synthetic write control hold and confirmed everything is working in the logs, started listening, finishing idle, speech analyzer starts, HUD renders, etcetera, etcetera. So what do I need to do?
Write control, testing. Oh, look at this.
Testing. Hey, can you hear me? Oh, oh, look at this.
Oh, that's a nice little touch. Just continue to test here. I just want to see how much you can pick up here and how fast you are if I just keep ranting like this.
I like the waveform too. It's it's a fancy it's a fancy little waveform. Okay.
Where'd it go? Where'd it go? Did it capture that?
Testing again. Can you hear me? Alright.
It's close, Claude. It picks up on my voice, I think.
Do you see this in the logs? It's just not inserting it where my cursor is. Please look at logs and fix.
Okay. The log show it's working further than you'd think. K.
Transcription and injection both succeeded. Then why is it not working? Let's see.
What does it say? That's the classic a x oh, yeah. Classic a x silent failure.
Return success on electron apps. Oh, interesting. Okay.
So it's a cursor thing when I'm trying to type in cursor. Alright. So it's fixing that now.
Alright. Rebuilt and armed. This should work.
Testing. Hello. Can you hear me?
Hey. Look at that. That was fast too.
Dang. That was really fast. Let's do some long rant here.
This is just a long rant here just to test how well this is working. Let's just keep talking. I do like this little thing in the on the right here where it just auto transcribes, even though it's kind of irrelevant because you can't even really see it.
But anyway, this is this is pretty good. Claude did a pretty good job for literally a one shot, more or less. Yeah.
There was a couple there was a couple, you know, permissions that need to be granted, and this a x, this classic a x error. Okay.
Let's see how fast. Woah. That is fast.
Holy. I bet that's faster than parakeet. That is crazy.
So I I haven't I haven't been using this new Mac OS model, so I might need to switch my own WhisperFlow clone to this. Testing this again, let's just see how fast this goes again. I'm gonna talk really fast and see what happens.
Dude, that's crazy. So holy crap. That actually worked way better than I even thought.
You can see how easy that was. Right? That was like one and a half turns.
We told Claude we wanted to create a WhisperFlow clone. We granted some permissions. We had one minor issue with just inserting into electron apps like cursor, and then now this thing is fully working.
We have a WhisperFlow clone that we can use, which by the way, I'll include as a repo in the link below if you wanna just take this and run with it. But I mean, pretty freaking cool. So what we're gonna do next is we're just going to show you how to download a local model just in case you wanna do this.
But it looks like this Mac, this new speech transcriber is very fast and works really well. And I'm just gonna tell Claude to do that with our new Murrur YouTube. So good job, Claude.
This looks great. It's very fast. We're using the Mac Mac OS 26 speech transcriber.
It's it's really good. I wanna do a side by side test with Parakeet somehow. I don't know how you can do this, but just figure out how to run a test.
Maybe just come up with a quick little dashboard and I'll just record something. We'll see how quick and accurate each of these are. I want you to explain and just show how we grab this parakeet model via Hugging Face too.
So okay, it's doing two things. The first is it's accessing this model via Hugging Face. I have the Hugging Face MCP server quick random aside that makes it easier to access and find these models.
But you might not have that. I I don't wanna get into setting up the MCP server, so I just said don't use the MCP server, so it's just using plain curl to access this model via Hugging Face's public API, so anyone can do this. I do highly recommend that you download the Hugging Face MCP server, but that is outside the scope of this demo.
This is gonna access that via Hugging Face and then just download this Parakeet model and then we're gonna do a side by side test. Okay. Cool.
So we have something built here. Cloud just needs to run this. What I'd expect so you can judge the results, Parakeet should win on raw RTF by wide margin.
Oh, interesting. Okay. We'll see about that.
Cool. Just build this. I don't know what else to say.
I don't know what it wants for me. Alright. So Claude just got done putting together this engine comparison.
We're gonna compare Mac speech transcriber versus Parakeet, the local model that we just downloaded from Hugging Face. We have this handy little window here, engine comparison, and all I need to do is just hold control and test to see how fast and accurate each of these are.
So I'm just gonna say, okay, just testing this record now with our Murmur YouTube Whisper Flow clone. I'm gonna talk for a minute here and kind of go on a rant. I may have to cut this down because this is kind of ridiculous to spend like thirty seconds to probably a minute talking.
Complete nonsense. Okay. Wow.
Jeez. Those are freaking Both of them are really fast. Okay.
So looks like Parakeet's a little bit faster. Zero point three two seconds, 0.58. Let's see what actually came out in the transcript.
Okay. Just testing this with Murmur YouTube, Whisper Flow clone. Yep.
That all looks alright. Okay. Interesting.
So Apple called it Cologne, which is wrong.
We won't be able to tell quickly you will actually pick up some testing this, and I honestly can't even believe. Let me test something else too. I also wanna see how well you pick up on words like Claude code.
For whatever reason, Claude code is very hard to get right. All of these speech to text models don't seem to get Claude code right. Let's see how well it does.
Parakeet's still faster. Do they get I can't get Claude code. Claude code.
Claude code. Alright. This is what we're gonna do.
We're just gonna go with Apple just because it's native to Mac OS. It's a little easier to set up. You don't need to go through the trouble of downloading a local model via Hugging Face.
It'll just be available pretty much out of the box in the GitHub repo below. But what I will do, because I know, of course, that means if you're a Windows user, this won't work. So I'll include a separate GitHub repo below that will use Parakeet and make our own version of WhisperFlow available to you.
I'll include instructions too on how to actually download that local model. And you'll be able to just copy the repo, give it to your agent, say, figure this out. Alright.
So that being said, our WhisperFlow clone is fully working. But the question is, how does this compare to WhisperFlow? Is WhisperFlow faster or better in any way?
Let's just quickly find that out. I'm gonna pull up WhisperFlow and compare it with our clone somehow and see how accurate and fast it is relative to our clone. Alright.
So Claude added something to our comparison engine where it's pulling from WhisperFlow's database. So I have WhisperFlow open here, and in our engine, if I open this comparison view, I can record all three now.
So it's gonna record Parakeet, it's gonna record MaxSpeechTranscribe, and it's gonna record WhisperFlow, and we're gonna see how quickly these transcribe and how accurate they are.
So let's go record all three. Okay. Testing this now with WhisperFlow, with Mac Transcribe, with Parakeet.
I'm just going to go on another rant. This is gonna be a quick rant than the last rant, but I'm just gonna keep talking so we can really test the speed of all three of these models, see if WhisperFlow is really that much better. At this point, it's gonna be so marginal that it doesn't really matter in any way.
Like, what is WhisperFlow's differentiator here? Let's see.
Alright. So it looks like Parakeet was the fastest. Zero point two seven seconds.
Apple was second, and then WhisperFlow was third. So okay. Let's look at these transcripts.
Just gonna go to the rant. This is gonna be quicker. See if WhisperFlow WhisperFlow got its own spelling correct.
Apple's missing some things like saying rent instead of rant. Apple got that question mark at the end, though. That's pretty good.
Alright. So there we go. Both of these models are officially faster than WhisperFlow, and our little clone is completely free, and you can see, at least in this example, that it is just as accurate in the actual text transcription.
Now the one caveat, and Claude calls this out here, is there is a little bit of latency because we're actually pinging WhisperFlow's database to get how long the transcription takes. So there's likely some latency, and so that's why that number is showing up much slower. But even if it was, so if we look at the numbers, even if it was a second slower, it would still take zero point nine one seconds versus zero point four eight versus 0.27.
And there probably is a way to be more methodical and get accurate numbers between WhisperFlow's model parakeet and MAC speech transcribe, but the point being, the difference in speed and accuracy is marginal at best. Now, we could stop right here and just have our little waveform HUD, but let's get this close to WhisperFlow and build a proper interface, a GUI as the pros call it.
That way we can start and close the app and just have it function very similar to WhisperFlow. So what I'm going to ask Claude to do is have this dictation history. We're not gonna do notetaker or anything like that.
And then we're gonna add a dictionary too and an ability to add words. So this is the prompt I'm gonna give to Claude. Basically, I'm saying I'll just open it up in this doc here.
I'll include this in the link below as well if you need it, but I'm just saying turn this into a proper Mac OS application, not a menu bar. The main window holds past transcriptions, start and stop, a dictionary, exactly what WhisperFlow contains.
By the way, to get this prompt, I just went to a separate Claude agent that was familiar with this project, and I said, what would be a good prompt to give to this agent? And then I also wanted it to lean into some kind of design direction, so I said, the direction is a nineteen eighties tape recorder. Claude just listed out some brands.
And then it just said, don't use neon vaporwave, synthwave, purple and pink gradients. I feel like these are just too overused. So I really wanna just like lean into some kind of design style and have it be kind of quirky and fun a little bit different than these other voice transcribers.
So I'll get Claude working on that. I'm guessing it will be built relatively quickly. Alright.
So Claude has built our Murmur YouTube interface. We've got the application, we've got transcription history, we've got dictionary. So let's open it up and check it out.
So here it is. It doesn't look bad. It's a little plain, but I guess it technically got the recorder kind of eighties feel.
Could've done a better job, but let's see if it actually works. So I'm just gonna hit record. Okay.
Testing this record now to see if our Murmur YouTube GUI is actually working. How is it going to transcribe GUI? Okay.
Didn't get GUI. Right? GUI?
But, okay, testing this record now. Murmur YouTube. Okay.
Well, there we go. We have transcription history. So let's go to dictionary.
I just said Claude adds some random phrases, so Claude code. Let's see if it'll pick up on Claude code. Okay.
Testing this now by saying the word Claude code. Let's see if it will actually correctly transcribe Claude code.
Boom. There we go. Cool.
Oh, and it even shows like a correction, corrected. Wow. That really does that very fast.
Like, corrects it in real time like that. K. Testing this Claude code.
Claude code. Boom. Just like that.
Alright. So we have dictionary. We can add different words here for the model to pick up on that transcription history.
I mean, that's basically all you need. Need anything fancy, and then you can just close this and have it running in your menu bar, and then just be recording stuff. So I could say, like, could open that Google Doc again, just say testing this again, our Murmur YouTube transcribing.
Let's see if this actually works in this Google Doc. Boom. There we go.
Beautiful. So that's gonna do it. That's how to clone WhisperFlow or really any of these voice dictation apps in just a couple prompts with Claude Code.
And don't get me wrong too, I have a lot of respect for WhisperFlow. They've built a great brand. I mean, their marketing's fantastic.
And if you want something that just works immediately, it is a great option. But if you're trying to save money, if you're trying to own your own data, if you're trying to customize this to your liking, you can see just how easy it is to do. Not to mention, if you don't wanna build any of this yourself, I'll leave the repo below.
Literally just copy it from GitHub, give it to Claude or Codex or whatever you use, and just say, this up on my own machine. So I appreciate you watching. There are so many apps like this, by the way, that could just be cloned in really just a few prompts with Claude code.
So if you have any ideas, if you have any apps that you use daily that you're paying a lot of money for, let me know and I can just put these together, make the GitHub repos available, and give you the ability to have fully custom software. Alright. I'll see you in the next one.
The Hook
The bait, then the rug-pull.
Wispr Flow just raised $280 million at a $2 billion valuation for push-to-talk dictation, a feature this creator says he rebuilt from scratch, in one Claude Code session, using nothing but Apple's free on-device speech model.
Named ideas worth stealing.
01:35list
Wispr Flow clone architecture (8-piece stack)
- Shell (the app container)
- HUD (recording indicator overlay)
- Hotkey (push-to-talk trigger)
- Audio (microphone capture)
- STT (Apple's native speech transcriber)
- STT-alt (Parakeet, a swappable backup engine)
- Cleanup (small model for punctuation)
- Injection (inserting text at the cursor)
The 8-part architecture Claude Code proposed after a single one-line prompt describing a Wispr Flow clone, each piece mapping to one concrete engineering decision.
Steal forany local-first macOS utility you want Claude Code to scaffold in one shot
08:14concept
Engine comparison dashboard
A small side-by-side testing tool Claude built on request that records the same sentence through multiple speech-to-text engines at once and reports latency and transcript accuracy for each.
Steal forbenchmarking any set of interchangeable AI models or APIs before picking one for a project
How they asked for the click.
VERBAL ASK
14:41link
“I'll leave the repo below... give it to Claude or Codex... and just say, set this up on my own machine.”
Low-pressure, resource-based CTA pointing to the GitHub repo and blog post rather than a subscribe or sale ask; ends by inviting viewers to suggest other paid apps to clone next.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
