Pinned Tweet
You can now write your own Rephrase styles: a name and one instruction.
The one in the video, TL;DR, is just "Three short bullets. Keep every number."
What would you put in yours?
2
173
Grok bot is literally the most advanced AI tool I ever used, everything is so smooth and so intuitive. It feels unstoppable
3
35
Your "Senior Engineer" title is the new "master scribe".
In the Middle Ages, copying a book took months.
So monasteries had a whole hierarchy: novices copied, experienced monks proofread, a master ran the scriptorium.
That pyramid held for one reason: copying was slow, and nothing could speed it up.
Then the printing press arrived.
Copying became almost free. The pyramid had no reason to exist anymore.
What mattered now was knowing what to print, and why.
Junior → senior → staff is our scriptorium.
Juniors to write, seniors to review, leads to coordinate. Because writing code took time.
AI is our printing press.
Hard skills barely set anyone apart anymore. What's left is curiosity, judgment, the ability to push back.
That used to be what got you to the top of the pyramid. Now it's the entry ticket.
So if you're a junior or a student today: where do you see yourself in 10 years?
1
1
76
Guys, now that AI is part of the daily workflow, how have your career goals and personal motivations shifted? What still makes the work feel meaningful and satisfying to you?
1
52
Today Claude asked me to leave the Mac alone during screenshot runs so it could work
Basically told me to get out 🤣
What a world we live in
2
52
Just joined a new company, which means a new team and onboarding onto a new codebase with Claude. I expected it to be easy. It's actually pretty challenging, I always feel like my understanding is a step behind Claude's, especially on deep harness topics
A speech-to-speech model will backchannel whatever your prompt says. Mine banned “mm-hmm” by name, and on my worst 30-minute take GPT-Live still said 44, 39 while the candidate was still talking. The candidate was me, it’s strange to get mm-hmm’d by your own product.
40
agree @poteto and articulation is the part people practice least. most prep is reading designs, but the round is about saying the tradeoff out loud while you draw. I'm the founder of thinkaloud.dev a voice mock where you talk through a design and get a report quoting what you said. still a beta with rough edges
i feel like swe interviews could probably collapse into just 2 technical rounds:
• system design (how well can you articulate your ideas, do you actually know how to engineer something)
• onsite project to build a real thing with agents (can you translate your intent into high quality outcomes)
everything else we used to ask is no longer necessary
1
34
Your voice agent doesn't cut people off because it's impatient. It's because you sent it context at the wrong moment.
Appended while the user talks, the realtime model spoke within 1.5 s 13 times in 34. Appended while it talks itself, 1 in 35.
2
4
138
Update: you don't pick a style anymore. Select text, ⌃⌥R, Enter, like before. In the background Jev from @typesafeai looks at what you wrote and where it's going, and picks the right tone in about 0.2s. On every plan, free included. Honestly the feature I wanted from day one.
The AI thing I use most is almost embarrassingly small. I write "I think we should maybe consider pushing the deadline" and I want it back without sounding like I'm apologizing for existing.
So I made it one shortcut on my Mac, and today it's out.
1
96
The AI thing I use most is almost embarrassingly small. I write "I think we should maybe consider pushing the deadline" and I want it back without sounding like I'm apologizing for existing.
So I made it one shortcut on my Mac, and today it's out.
🤖 Made with AI
Before, it was a Raycast command: copy, open Raycast, type rephrase, enter, go back, paste. The AI part was fine, the rest was the problem. Now you select text in any Mac app, press ⌃⌥R, then Enter. 20 free a month, Pro is $29.99/yr ($2.50/mo). rephrase.studio
🤖 Made with AI
50
The TTS was hallucinating in German, Spanish and Portuguese and it was not the model. It was ElevenLabs streaming with auto_mode on partial sentences. We feed it whole words now, and the hallucinations dropped roughly 5x.
53
Victor Billaud retweeted
A voice agent that can't see images, running a whiteboard interview. The canvas reaches it as plain text, no model writes it. Jev reads what I draw and what I say, 39 typed questions in under half a second, and code picks the next question. The voice never hears a judgment.
My voice interviewer can't see the whiteboard; GPT-Live takes no images. The box I named "Fastly" is read as a CDN at 1.00, and my note "sync or async transcode?" as a question I still owe. Jev answers dozen of questions like that in under half a second. The voice never hears one.
The first try was a compact grammar with IDs and a 3x3 grid for positions. It broke on one stray box far from the rest; the grid squashed everything else into one cell. And a realistic canvas of 14 elements was about 90 tokens anyway, so saving tokens was solving nothing. In the end, it's plain English from a deterministic renderer, one fixed sentence per element type: "Client" → "App server" → "Queue". No model writes it, so a replay gives the same text byte for byte.
"This one" was the other hard part; you can't guess it from geometry. The browser knows what I just clicked, so every snapshot starts with a Focus line and the voice is told to trust it.
Then @typesafeai jev. It never speaks; it answers small, typed questions, and each answer is a probability. What is this label? Is this note an open question? Did I talk about this box? Did I answer my own note? We wait 3 seconds of quiet so it doesn't judge half a sentence, then one request carries all of them. On my upload line, that's dozen of questions, back in 150 to 200 ms.
None of that goes to the voice. Because a wrong guess said out loud puts words in the candidate's mouth. The judgments go to plain code that picks the next question and which box to name. The voice gets the labels, never the reading.
It's not perfect yet, but wow!










