ege oz (@egeozin) header
ege oz (@egeozin) profile picture

building autonomous companies at https://nitter.gq/t.co/rC6agOfUFY, previously @odastudioapp (yc w20) @MIT

San Francisco, CA
Joined April 2011
went to mit. spoke to a capybara on ayahuasca. literally floated during jhana. deadlifted 3x the weight of my brain. clenched and unclenched 78 times. finally found my purpose. i will market your saas. we're open-sourcing tin, a marketing system for early stage founders that you can use from claude code, codex or cursor. our first approach was basically codex running in the backend, with tools, skills and a lot of freedom. some of it worked, but we also got a lot of slop: plans that kept getting rewritten, repeated tasks, drafts that never went anywhere. figuring out which part to fix was its own special kind of hell. so we rebuilt tin around opinionated workflows. for example: - audit your website, use the findings to plan content, write a draft, revise it with your feedback and, with github connected, open a pr after approval. - assess whether google ads make sense for your business, launch a campaign after your approval, then monitor search terms, pause wasteful keywords and bring budget changes back to you. - sign up for your product with a fresh account, verify the email, actually use its features and report where things break. each step has a defined job, relevant project context and an output the next step can use. the harness includes workflow evals, feedback loops, retries and scheduled runs, so we can test changes and keep improving the workflows individually. what i want this to help with is trying different marketing approaches, seeing where the signal is and deciding what to double down on. yeah, call it vibe marketing. connect your coding agent through mcp. you can read and change the workflow instructions, self-host or use our hosted version. there's plenty to improve. please try it on your project and tell us what's missing.
🤖 Made with AI
17
4
3
37
2,655
sold all my nvidia (hardware). went all in on docusign (wetware)
1
2
199
ege oz retweeted
Qrambo: an operating system for operations, built on top of the systems you already have. Demo video link in the thread... 🧵🧵🔗🔗
1
1
3
149
i think this debug view is cooler than the internal canvas tool we are building to render video and javascript animations. @stephenhaney live code layers in paper, wen?
4
195
hey i’m ege and i’m building a sandbox marketplace called kindergarden
2
7
330
Revamped an old project of mine, Creative Sanctuary Deck: 52 cards for when you’ve been staring at your work too long and your brain has stopped being useful. Some are encouraging, some are weird.
1
2
64
yo i made a site that lists games whose code is public: decomps, static recomps, reimplementations and source releases. 4,573 games so far, each linked to its repo: unbaked.games
1
2
10
593
stuff i do at 2 am, violating 2 IPs, but worth it
1
1
8
321
fix your claude code designs with this simple trick: "no eyebrows anywhere"
1
5
305
we soft-launched tin computer last week and added $250k ARR most of the new clients had no time to try another ai tool and set it up for their own case. they asked us to make growth happen for them, so that's what we're doing if you're selling an ai tool, try selling the whole service around it
1
2
1
14
617
yup i'm using this: don't be a compute bag!
i should start a group for people who aren't literally psychotic or completely one-shot some kind of group where only humans who don't have the most brittle and simplistic weltbild, their picture of the world, are allowed in bonus point if youre broad enough to have an interests ranging from branches of theoretical computer science (not just being a compute bag who can write leetcode hards), cognitive science, developmental biology, a penchant for proper abstraction, theory of computation, and actual hobbies that have nothing to do with this (and where the hobby's goal is not about you being a compute bag if possible!) and then we just close off entirely and just collectively block all inputs from retards like this - maybe we have a few tweeners who still take in inputs and report and that's it
2
396
if you want to train custom voice models for your use case, generate synthetic data for your domain with VibeVoice and fine-tune a duplex model on it. we simplified it for you →
Thinking Machines and OpenAI have released live duplex conversational models recently. But what's going on with open source models? As Emotion Machine we're releasing our PersonaPlex finetuning pipeline - with synthetic data generation, a puppeteer and bot2bot evaluation.
5
674
I built a public map of every reported whale stranding in the San Francisco Bay Area since 2005: 216 records from NOAA's Marine Mammal Stranding Network database. I'm not a marine scientist, but I've followed the strandings for years and wanted a way to see them all together.
1
1
2
233
a lot of people have started using with.md, so here are this week’s updates: - you can now create and edit markdown files from within your claude code & codex sessions. your agents can edit the files they created too! install the skill by instructing your agent to read `with.md/skill` & `with.md/skill/md`. we don't recommend installing the skill via npx, but if you prioritize convenience over security, you can install it via `npx skills add emotion-machine-org/with-md`. - you can now convert urls to markdown. try opening a blog/docs/article by prepending `with.md/` at the beginning. example: with.md/{your-url-here} - with.md now has a cursor/vscode/openvsx extension! install it and click on any markdown in your repo to view it with with.md again: with.md is open source. if you have ideas for extending, securing, or improving it, feel free to open a pr on github or reach out to us.
we’re open-sourcing our internal markdown editing + collaboration tool. built with @demegire1 and @emresarbak our team is spending more time planning on markdown files than coding. you probably are too. why is it so hard to collaborate on markdown files? i don't know. we just couldn't find anything that made it easy to share and work on the same file. sooo we created a new markdown file to plan a markdown editing + collaboration tool and vibe-coded with.md with.md lets you: - drag-and-drop your internal .md docs (no sign-in required) or sign in with github and access all the md files in your repo to organize, edit, and share repo-scoped markdown - share them with teammates to edit together (or get sign-off) - then, feed them back into your claude code / opencode / etc. - we also made shares easily parseable by agents. we’re planning to add more workflows like real-time editing + attribution for claude code agents, arxiv/docling integration, turning urls into markdown, etc. but that’s down the line. we weren’t avid markdown users, we mostly used google docs for company-wide stuff. but as we started using more and more terminal agents to edit, improve, and then straight-up copy-paste, we ended up sharing more and more markdown. and for whatever goddam reason, rendering markdown in cursor/vscode is a pretty clunky experience. then, if you share a markdown on slack, the other person has to download it, open it in their own vscode/cursor/sublime, and get the same shitty md viewing experience. we didn’t like that, so we built something that works for us. realtime cursors, anchored comments, diff view (thanks @pierrecomputer) -> the stuff you’d expect, but somehow doesn’t exist in one place for .md files. make no mistake: our understanding of wysiwyg editors and how crdts work is still pretty limited. but after reading the crdt docs on @ekzhang1’s reading group site, @geoffreylitt and the ink & switch crew’s peritext post, and talking through a primitive v1 design, the rest was straightforward: write a thorough plan in md, turn key resources and docs into md, feed it into codex 5.3, and get a working demo. after that, we polished the rough edges, iterated on the md rendering design system, fixed bugs, and tried to make the whole thing reasonably secure with a few extra human/agent passes. you’d think doc editors are “gg ez” they’re not. still, it was encouraging that the tool we built for ourselves (humans) was necessary to get to something that feels worth sharing on twitter. there are still a lot of bugs, we’ll try to squash them as we keep updating the tool. this is not a replacement for your @obsdmd , gdocs, or @NotionHQ . for our specific use case, we couldn’t bootstrap something more practical with existing tools. but yeah, if you have other ideas & tools, please share. again: this is 100% vibe-coded, but it works. we’re open-sourcing it to make it better. the github issues section is waiting for your snarky feedback.
3
1
1
9
999
we’re open-sourcing our internal markdown editing + collaboration tool. built with @demegire1 and @emresarbak our team is spending more time planning on markdown files than coding. you probably are too. why is it so hard to collaborate on markdown files? i don't know. we just couldn't find anything that made it easy to share and work on the same file. sooo we created a new markdown file to plan a markdown editing + collaboration tool and vibe-coded with.md with.md lets you: - drag-and-drop your internal .md docs (no sign-in required) or sign in with github and access all the md files in your repo to organize, edit, and share repo-scoped markdown - share them with teammates to edit together (or get sign-off) - then, feed them back into your claude code / opencode / etc. - we also made shares easily parseable by agents. we’re planning to add more workflows like real-time editing + attribution for claude code agents, arxiv/docling integration, turning urls into markdown, etc. but that’s down the line. we weren’t avid markdown users, we mostly used google docs for company-wide stuff. but as we started using more and more terminal agents to edit, improve, and then straight-up copy-paste, we ended up sharing more and more markdown. and for whatever goddam reason, rendering markdown in cursor/vscode is a pretty clunky experience. then, if you share a markdown on slack, the other person has to download it, open it in their own vscode/cursor/sublime, and get the same shitty md viewing experience. we didn’t like that, so we built something that works for us. realtime cursors, anchored comments, diff view (thanks @pierrecomputer) -> the stuff you’d expect, but somehow doesn’t exist in one place for .md files. make no mistake: our understanding of wysiwyg editors and how crdts work is still pretty limited. but after reading the crdt docs on @ekzhang1’s reading group site, @geoffreylitt and the ink & switch crew’s peritext post, and talking through a primitive v1 design, the rest was straightforward: write a thorough plan in md, turn key resources and docs into md, feed it into codex 5.3, and get a working demo. after that, we polished the rough edges, iterated on the md rendering design system, fixed bugs, and tried to make the whole thing reasonably secure with a few extra human/agent passes. you’d think doc editors are “gg ez” they’re not. still, it was encouraging that the tool we built for ourselves (humans) was necessary to get to something that feels worth sharing on twitter. there are still a lot of bugs, we’ll try to squash them as we keep updating the tool. this is not a replacement for your @obsdmd , gdocs, or @NotionHQ . for our specific use case, we couldn’t bootstrap something more practical with existing tools. but yeah, if you have other ideas & tools, please share. again: this is 100% vibe-coded, but it works. we’re open-sourcing it to make it better. the github issues section is waiting for your snarky feedback.
12
7
8
143
27,797
equinotes
Need a trading floor for SAFE notes at the Pacific Coast Stock Exchange
554
this should not be underhyped
Something new I've been hacking on: draw directly on your interface and leave annotations that agents can see and understand. Opens up a really natural/freeform way to iterate.
433