ve been using Replit for about a year and recently landed on a workflow that’s significantly sped up how I build more complex apps. The big unlock for me was combining voice input (Wispr Flow) with ChatGPT and Replit’s Planning + Build agents. I talk about 3× faster than I type, so being able to speak full context, assumptions, and edge cases upfront has made a huge difference in clarity and speed.
My general flow is: I talk through an idea using Wispr Flow, then pressure-test it with ChatGPT. I have ChatGPT challenge assumptions, suggest best practices, and think like a senior engineer or architect until the idea is solid. From there, I generate a high-level plan and a short PRD, feed that into Replit’s Planning mode, answer its questions, and then send the plan back to ChatGPT to critique it. I intentionally put the two AIs against each other until neither can meaningfully improve the plan—then I greenlight it and let Replit build.
This has let me one-shot large chunks of advanced functionality with minimal cleanup afterward, even for fairly complex systems (I’m currently doing this on an email marketing system with real-time ecommerce data sync and AI-generated content).
Workflow summary
Voice-first ideation with Wispr Flow
I talk through the full idea instead of typing. This includes the problem, goals, constraints, assumptions, edge cases, and things I’m unsure about. Because I speak much faster than I type, I can front-load a lot of context into a single prompt, which leads to better downstream results.
Idea validation and pressure-testing with ChatGPT
I go back and forth with ChatGPT to challenge assumptions and confirm best practices. I ask it to think from multiple perspectives (senior developer, architect, product, marketing) until it clearly understands what I’m building and why.
High-level architecture and step breakdown
Once the idea is solid, I have ChatGPT produce a high-level system summary and a step-by-step implementation outline. From that outline, I generate short PRDs for the first few major steps of the application.
Replit Planning mode for execution planning
I take the first PRD and paste it into Replit Agent’s Planning mode, asking it to create a detailed plan and to ask clarifying questions before starting. I answer what I can directly, and anything unclear I bounce back to ChatGPT for help.
Cross-checking plans (AI vs AI)
After Replit produces a plan, I give that plan back to ChatGPT and ask if it agrees with the approach, whether it follows best practices, and what it would change. I iterate until neither AI can meaningfully critique the plan anymore.
Build, test, and iterate in Replit
Once the plan is locked, I move Replit into Build mode. After it finishes, I ask for a test checklist, run through it myself, and fix small issues using Fast mode. Then I move on to the next PRD and repeat the process.
If you want to try the same voice-first workflow, this is the tool I’m using for input:
Hope it helps its really sped up my production of complex apps! Happy vibe coding and if you want to check out our vibe coding co-op we are starting you can join the waitlist here - http://joinvibehub.com/ we’ll launch once we have a sufficient amount of interest!
I use Perplexity’s Comet browser with its Assistant sidebar in the Replit IDE to do similarly. It’s as close to an in-line workflow as I’ve experienced.
I admit I haven’t tried voice input for Replit, but for good reason. Because I witter on at 200wpm, and know I would be constantly changing what I’d said.
For me, typing is still the overall fastest and most efficient.
I type relatively quickly, and then I go back and re-read and modify until I am happy with my prompt. The first version would always have been a mistake and caused the agent to go horribly wrong - and the same would go for any voice input.
Nice. I do need to start using Wispr Flow more (wish it worked on Android). Often times I just prefer typing because I flow better that way. My typing is pretty fast.
I sometimes do a version of Agent ← –> ChatGPT for big changes, but usually I just iterate on the idea with Replit Agent and concentrate more on the building and iterating because I don’t want to be overly prescriptive so that I leave some room for pleasant surprises (Agent builds a feature in a way better than I expected / could have thought of myself).
During iteration I ask it to play specific roles which I think are most helpful for the task at hand (CTO, CMO, Designer, etc), ask it to check online for b2c best practice for the topic at hand, and also ask it to check with the Architect before and after implementing. My way can be more expensive, but I like the fast iterative method where I provide general direction, do some refining, and let it flex its creativity a little.
That’s the thing about AI voice, it’s so easy to just talk to it, and tell it to remove things, or rewrite. It’s not at all like the dictation of years past. But I hear you. I also type at lightning speed, but it has certainly helped me be more productive.
I don’t mean to re-bump this… but as a student with disabilities who went through the public school system… I’ve been forced to use the worst of the worst.
Y’all just blew my mind with wispr flow
If anyone here remembers the Kurzweil program and other public tools (yes, before AI). You’ll know how these programs add several hours of additional learning, staying up to date with their software changes and can often cause catastrophic failure.
I mean if you think writing code is “so last year”…
Imagine going into a computer science exam, or university level mathematics that forces you to produce pin point accurate syntax and show all work, and the tools they provide you are:
&
Equatio for typing complex math equations.
So, basically, you need to learn how to write using LaTeX, and then top it off with producing pristine Java Syntax using the sometimes working Dragonfly speech to text workstation…
You better be able to speak programming syntax basically…
They act like it’s NBD btw just give you an extra hour, no formal hands on training, no classes in LaTeX lol
Just have fun, and good luck! This is real life bud.
I use some tools like:
For reading text aloud.
I’m incredibly thankful I found this in the space from you and rayymarr
Such a powerful tool
If only low budget public schools could be given access and students with accommodations could start utilizing these tools instead of the truly antiquated ones they end up failing entire courses because of…
What’s funny now is the challenge has almost flipped. Back then the tools slowed you down so much it took forever just to get work done. Now with things like Replit, voice input, and all the AI tools, you can get way more done in a day or a week than was even possible a few years ago.
In some ways it’s still hard, just in a different way, because there are so many possibilities and things moving fast. But overall it’s a huge step forward. And it’s only going to keep getting more agentic, more automated, and more streamlined from here.
Really appreciate you sharing your experience. It’s good perspective for people who haven’t had to deal with those older systems.
Honestly, at the risk of tarnishing, my credibility I’ll admit having to rely on these outdated systems is in part what caused me to drop out and fail one of my semesters.
Unfortunately, these old outdated systems can actually exasperate or add unintended barriers while trying to remove other barriers.
Eventually, it becomes self-defeating and as the student you have to decide if you’re going to spend the money to purchase tutoring outside resources and top-tier products, or use the schools provided system which will require dedicating almost an entire semester to learning before using them in practice.
Otherwise you have to struggle and not use anything or a mix of whatever you can on the fly.
However, I’m typing this or speaking this with my Siri speech to text and I gotta say it has come in handy over the years. It’s far superior than any of the paid university subscriptions, which is wild. I guess that’s apple for you though.
Although, if you have a weird accent, like I do sometimes stutter and mumble. Siri can take words like mute and moot and not know the difference.
I use WisprFlow with replit now and it’s great! But I really keep wanting the Replit agent to speak back to me with audio. Has WisprFlow added that yet?
This would be incredible and please don’t charge me for it.
Just read it aloud and consider it standard basic accessibility accommodations lol
Read the codebase.
An agent that somehow is able to read code aloud that’s not awkward… would… be… a… game changer.
Random side question:
Has anyone else heard their AI reader do something oddly human, like burp or stutter, roll over their words like their drunk, or something just odd for a AI but make the experience more authentic lol
Not a drop since 9/22/2021 around 3 AM Eastern Standard Time ROFL
No, Eric, this is sober me.
Hahaha
I’m just sifting through and being a dork
I really did appreciate finding wispr…
Hardware suggestions would be dope as I’m not into that and would not know where to start on microphones.
I’m developing tools and gathering resources, and become proficient to expert level with the tools now before I get back into classes so when the time comes to sit down and take the exam or do the project I’m not spending the entire time fiddling with the settings and just getting the stupid things to function…
Really, all jokes aside… Try taking the exams required for formal degrees with the provided accommodation tools provided by state or local universities…
I’m talking about calculus exams where you have to show work and one wrong step will ruin the whole thing.
If you mess up even a single step, or miss remember a single formulae (ahem quadratic formula) then poof. There goes all your points.
Force shoving Java syntax out either using speech to text or through memorizing the syntax keystrokes.
Reading multiple choice questions and their provided options aloud and hopes it doesn’t select options for you and skip the next question lol
It’s mastering tools like this that will speed up development and make long term educational venture sustainable for me lmfao
probably need to get over that
