OpenAI launched Prism in January 2026, and it's already shutting down on January 1, 2027.
I guess OpenAI tried to do too much. I think many products like Prism are better left to the community to build.
If you want to work on LaTeX and Typst locally with the agent you like, such as Codex or Claude Code, give suzuri.ai/ a try
#LaTeX #Typst #AcademicWriting #ClaudeCode #Codex #AIAgents #Prism #Suzuri
I have been doing all 5 for PaperFox.ai and I think I am going to develop a course to train our students to do that too.
As engineering, product, design, DS, etc. melt into a new kind of role, I was reflecting on what roles might look like in the future. For example, when I look at the Claude Code team I see what I think is five archetypes:
1. Prototyper: comes up with brand new ideas; churns out many ideas, most of which don't ship
2. Builder: quickly turns a prototype/idea into production-grade product/infra
3. Sweeper: cleans up the UI, simplifies the code and system, unships, optimizes performance
4. Grower: takes a product that has been built and iterates on it to improve Product-Market Fit
5. Maintainer: owns a mature system to make it secure, reliable, fast, and efficient as it scales
Many people span across 2 roles, and sometimes 3 roles. I also notice that these roles are not really tied to job function -- eg. across Anthropic, some designers match category 1, some 2, some 3; same for engineers, PM, DS.
A healthy team needs a mix of these, depending on the product:
- A product that is new and pre-PMF needs people that are strong at 1+2+3
- A product that is growing and has found PMF needs 2+3+4 and some 5
- A product that has strong PMF needs 3+4+5 and some 2
Maybe product roles of the future will look more like this, and less like the domain-specific roles of today?
I recorded this clip ("Fox" by Xiaoli Wan) at a live show in Chengdu, China this weekend. It's another good example of how I use Claude Code: I added bilingual subtitles in under 10 minutes with a few simple prompts and no video editing software, a job that could have taken hours by hand. Here's how:
1. Transcribe: Claude Code extracted the audio from the 58-second clip and sent it to OpenAI Whisper, which returned a timestamp for every word.
2. First draft (v1): It split the words into subtitle lines at the pauses, then had GPT-4.1 clean up the Chinese and translate it into English.
3. Fix the lyrics (v2): Whisper misheard much of the singing, so I gave Claude Code the official lyrics. It swapped in the correct text while keeping Whisper's timestamps to place each line.
4. Fix the translation (v3): GPT had translated slang and idioms literally. When I pointed out one example, Claude Code immediately understood the problem and fixed every other literal translation as well.
5. Burn in: My ffmpeg build can't render text onto video, so Claude Code drew each subtitle as an image with Python and overlaid it on the footage. The result: a 1080p MP4 plus a separate .srt subtitle file.
#ClaudeCode #AI #AICoding #OpenAI #Whisper #Subtitles #VideoEditing #LiveMusic #Chengdu
I asked Claude Code to cut a one-minute highlight from my tennis practice video, analyze my problems, and suggest improvements. The suggestions were very good.
I also asked it to make an image showing how it did the work (below). It turned the 4-minute phone video into contact sheets, listened for the "pop" of each hit to find the busiest rallies, broke down key strokes frame by frame, and stitched the highlight reel with ffmpeg. Quite amazing.
#AI #ClaudeCode #Tennis #SportsAnalytics #ComputerVision #VideoAnalysis #ffmpeg #AIAgents
Harry Wang retweeted
Domain search no longer requires auth.
▪︎ Search by keyword with the Vercel CLI
▪︎ Batch-check up to 200 domains via API
𝚟𝚎𝚛𝚌𝚎𝚕 𝚍𝚘𝚖𝚊𝚒𝚗𝚜 𝚜𝚎𝚊𝚛𝚌𝚑 𝚊𝚌𝚖𝚎𝚜𝚒𝚝𝚎
vercel.com/changelog/search-…
I made a narrated explainer video for PaperFox.ai with Claude Opus 5.5. No video editor, no HyperFrames, no Remotion. The voiceover cost about one cent.
I built a Claude Code skill that turns any docs page into a short video:
• Claude reads the page and writes a 30–60 second script, one idea per scene
• Each scene is an HTML page, rendered frame by frame with Playwright
• OpenAI's gpt-4o-mini-tts reads the script in the "marin" voice
• The background music is generated in code, so there's no licensing to worry about
• ffmpeg puts it all together, ducks the music under the voice, and adds captions
gpt-4o-mini-tts costs about $0.015 per minute of audio, so voicing this video cost about a cent.
Imagine what a video like this would have cost in time and money without AI.
Watch the short video on how phases work: paperfox.ai/docs/track-chair…
#AI #GenerativeAI #ClaudeCode #OpenAI #TextToSpeech #VideoProduction #Automation #Playwright #Documentation #TechnicalWriting #AcademicConferences #PaperFox
Harry Wang retweeted
excuse me I must correct the record on this slander
🤖 Made with AI
You guys think @alexandr_wang is underdressed but he's literally flexing on you like crazy lol
There are levels to this
Readers added context they thought people might want to know
None of these clothing items exists. Hermes doesn’t make hunter t-shirts, Loro Piana doesn’t make zipper sweatpants, and Ferragamo doesn’t make high-end crocs. us.loropiana.com/en/ ferragamo.com/shop/us/en/men…
Suzuri keeps getting more community suggestions, bug reports, and code contributions. Creating something people want, and then working with those people to make it better, is a very good feeling.
I'm now working on Zotero integration and porting over a number of Typora features. More on that soon.
Bug reports, feature requests, and code contributions are all very welcome: github.com/harrywang/suzuri
#Zed #Obsidian #Markdown #Latex #Typst #Zotero #Typora #ScientificWriting #LocalFirst #KnowledgeManagement #OpenSource #DeveloperTools #BuildInPublic
A question for one-person company (OPC) founders selling to businesses: how do you handle business continuity and risk questions in sales conversations? Consumers rarely ask, but organizations need confidence that your company will be around for the next few years and that you're not a single point of failure. I believe an AI-native OPC can build better systems than many larger teams, yet earning that trust remains a key challenge. How have you addressed it? @marclou @levelsio
The same day CyberCab was released, I tried two Chinese Robotaxi services in Shenzhen — Pony.ai and Apollo Go - both are impressive, especially given how much more complicated the road conditions are than in San Francisco, where I tried Waymo - see in the video how motorcycles go the wrong way and people cross wherever they like 😂 The Apollo Go car even has massage seats, and you can voice control pretty much everything, like enter last 2 digits of phone number, temp control, play music…Very cool. #cybercab #robotaxi #ponyai #apollogo #waymo
Funny thing — I found someone on Reddit looking for a tool exactly like Suzuri.ai I built, so I left a comment. He replied harshly saying that my comment was a spam ad, partially based on my user id.
I get it — there's so much spam and so many ads flying around on Reddit. But mine wasn't one: I've been on Reddit for years and posted things that got thousands of upvotes. Reddit just assigns you a random username if you sign up with Gmail, and it won't let you change it.
Initially, I didn't know why Reddit doesn't allow ID changes and got frustrated, but I ended up actually liking my ID — especially the 404 part that only developers understand. I never changed my random Zorro babysuit avatar either — it's cute too.
#Reddit #BuildInPublic #Suzuri #OpenSource
Yesterday I saw a popular Reddit post making a point I keep coming back to: when AI lets everyone build the same thing you're building, distribution becomes much, much harder than building (reddit.com/r/ClaudeAI/commen…).
This is the distribution era, and I shared a good article on it a while back: gtmnow.com/the-distribution-…
When anyone can build the tool, building it is no longer the moat — your reputation, your profile, and your audience are.
So alongside the tool, build two other things: a portfolio of what you've shipped, and an audience on social media and in the communities you actually care about. Here's my portfolio: harrywang.me/projects/produc…
That's your distribution, and it takes much longer to build than the tool does. Start early.
#AI #BuildInPublic #Distribution #IndieHackers #DeveloperTools #GTM #PersonalBrand #AICoding
Since I released Suzuri a week ago, we've merged our first community contribution — huge thanks to David Turnbull for fixing a GPU memory leak in the PDF viewer that hit live Typst and LaTeX previews hardest — along with a stack of features and bug fixes: editable frontmatter properties, a source mode toggle, LaTeX math in markdown, and Jupyter notebooks that keep their outputs.
Check out the changelog for the latest updates: suzuri.ai/log
I've been running my teaching and research in Suzuri for a while now — local-first, markdown-driven, working alongside agents — and the productivity gains have been real.
If that's how you work too, I'd love your feedback: try it and tell me what breaks or what's missing. And if the approach resonates, contributions are very welcome: github.com/harrywang/suzuri
#Zed #Obsidian #Markdown #LaTeX #Typst #ScientificWriting #LocalFirst #KnowledgeManagement #OpenSource #DeveloperTools #BuildInPublic
I usually have a bunch of Claude Code sessions running at once across different projects — and I kept sending prompts to the wrong one once in a while.
/rename helped but I just noticed /color. Now the routine is: rename and color a session when it goes active, and switch the color back to default once the PR is merged and the worktree is cleaned up. Pretty happy with this improvement.
#ClaudeCode #AI #AICoding #AgenticAI #DeveloperTools #Productivity #Workflow #BuildInPublic
If you like Zed and Obsidian, and you write LaTeX or Typst for your papers, check out suzuri.ai — a fork of Zed I maintain for markdown and scientific writing - open source with GPL-3.0.
I've been dogfooding it daily for a while now and really enjoy it — any feedback is highly appreciated.
The story behind the project and the name: suzuri.ai/about
Hope it's useful. Cheers!
#Zed #Obsidian #Markdown #LaTeX #Typst #PDF #ScientificWriting #OpenSource #DeveloperTools #BuildInPublic
We just shipped LaTeX support in PaperFox.ai — looking good.
#PaperFox #LaTeX #PeerReview #ConferenceManagement #AcademicResearch #BuildInPublic
Claude Code told me yesterday auto mode is now the default although I've been running it that way for a while — just switching modes manually every session.
I still remember the first time I ran it in YOLO mode back in January. I dug out an old, isolated MacBook just to run it on — that's how much I trusted it.
Six months later, I trust Claude Code more than I trust most human developers.
The trust curve moves fast.
#ClaudeCode #AI #AICoding #AgenticAI #DeveloperTools #BuildInPublic #HCI #Trust
























