Valentin Yeo

I let AI agents build my startup and post what breaks. Founder @hypertasks or - CRO at

[Valentin Yeo @valentinyeo]

Used the @felixrieseberg prompt to fix my CI and lots of learnings from @poteto and using pstack etc.

Shipped 500 PRs last week. I have adpoted the decent verification + fast rollback approach rather than trying to rule out bugs go to live this also helped to ship more.

I have 2 agent farms one for https://t.co/gJAqIdKEyu one for all my opensource stuff https://t.co/DvAgcRevjW which is mostly saas software that i attempted to replace with my own versions but adapated to my preferred workflow.

Trying to make these projects more accessible and useable for others down the road maybe... Not sure. Software is hard :)

[Valentin Yeo @valentinyeo]

@ItsAbuzar_ QA agents and bug reports from users as well as using it myself all day and then just turning the flag off haha. Not very professional but good enough.

[Valentin Yeo @valentinyeo]

It's so important for your product owner workflow to fix your CI and reduce the time to live as much as possible. For me its now 10 minutes for 90% of the tickets for https://t.co/gJAqIdKEyu with 100s of PRs every day.

In past iterations of my CI setup i had to wait hours for changes to appear. Might as well work with humans then 😬

Just kidding but my point is if the change on the live app takes a long time you lose creative momentum and work on other stuff instead of iterating on a feature until its right.

It's the only way you can make the live app also your iteration tool.

I recently re-introduced @figma into my workflow for complex UX features i am not sure of as I learned that the MCP is really capable now.

Last time i tested the figma mcp many months ago it was restricted and could not do everything I wanted.

For UX i am really not sure of this is faster than doing it live of course.

I wrote about it here https://t.co/ijWHLzR5Yr

[Valentin Yeo @valentinyeo]

Claude deadlocked my CI, and I can't code.

Four things got me past it. Big credit to @felixrieseberg, @poteto, @owengretzinger and @o_kwasniewski. https://t.co/7Dp8UWUQzc

[Valentin Yeo @valentinyeo]

This is what i wanna spend my Fable tokens on 👀

[陈成 @chenchengpro]

Prompts to use with Fable 5.2 to improve your codebase.

1/ Full health check. from Alex Albert，Anthropic

Read this entire project and find everything that's broken, inconsistent, inefficient, or no longer needed. Fix what you're sure about and list the rest for me to review.

2/ Simplify by cutting the superfluous. from Anthony Riera

I want you to audit our codebase, is anything here unnecessary, overly complicated in our
current setup, or based on weak assumptions? Challenge them. #1 What can be deleted entirely? #2 What can be simplified now that unnecessary pieces are gone? Then make the changes. Prefer deleting over simplifying, simplifying over optimizing, and optimizing over automating. This product is being used by thousands of customers, do not cut any corners.

3/ Find the implicit state machine. from Steve Ruiz and Kun Chen and RicardoSawir

audit and find implicit state machines in this <codebase/directory/module> that would drive
concrete improvements if made explicit. fewer words, big visuals

or with consider for UX/DX/AX.

Find all of the implicit state machines in my app (there are probably a ton). Make the best decision that will elevate all three: the user experience (UX), developer experience (DX), and agent experience (AX) always, without breaking anything. Also, make state machines explicit.

4/ Review agent configuration. from WquGuru & Vox

Review my entire agent setup: AGENTS.md, CLAUDE.md, skills, commands, subagents, hooks, settings/permissions, and MCP servers. Check for gaps in guardrails and whether it follows current best practices for Fable 5.2.

5/ shadcn/improve

run https://t.co/qoRSy4zk73 for my codebase.

[Valentin Yeo @valentinyeo]

I am now excited about grok model releases :D

[Elon Musk @elonmusk]

We have dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++, which took some extra time.

Training (aka pre-training) is done. Lot of cleanup was needed on the SFT data to get that mostly right. Now, 4.8 is in RL and the evals are starting to climb. We will do a soft rollout to Bot users later this week and see how it goes.

Grok 4.8 will bear no resemblance to 4.7 (thankfully). It is not yet incredible, but, this time, the foundations are built right, so I expect it to evolve to be something special.

[Valentin Yeo @valentinyeo]

If your 'software factory' or system is setup the right way most agents can deliver 'the right fix/ux' on first try. That's why releasing to live with a flag to turn it off fast is probably justifiable for some projects. Obviously if you serve to millions of customers that rely on it don't take that approach. But as a solo operator I think it has has to be considered.

[dax @thdxr]

it's crazy to me people still argue about subagents

most of our team tells opencode to do everything in subagents

it's also how we're drastically cutting costs and still getting reviews from expensive models

[Valentin Yeo @valentinyeo]

@poteto @maid_crypto funny 😂

[Valentin Yeo @valentinyeo]

@blud872 @felixrieseberg @poteto @owengretzinger @o_kwasniewski yeah outside claude code session basically took the fix over

[Valentin Yeo @valentinyeo]

Claude deadlocked my CI, and I can't code.

Four things got me past it. Big credit to @felixrieseberg, @poteto, @owengretzinger and @o_kwasniewski. https://t.co/7Dp8UWUQzc

[Valentin Yeo @valentinyeo]

@akshit_io Also not 100% sure. It would be good for the multi tasking, flow state aspect of working with agents.

One thing that serioulsy disrupted that is the chief of staff concept. If you only talk to 1 session most of the time an inbox becomes futile.

[Valentin Yeo @valentinyeo]

One view i added is an inbox view ... Inbox zero I believe is on the horizoin for agentic work.

Just like email give the agent a long running instruction and it comes back to your inbox when it's done.

[Vercel Developers @vercel_dev]

DeepSeek V4.1 Flash 1017 Ultrafast is now available on AI Gateway. It generates over 1,000 tokens per second.

Use 𝚍𝚎𝚎𝚙𝚜𝚎𝚎𝚔/𝚍𝚎𝚎𝚙𝚜𝚎𝚎𝚔-𝚟𝟺.𝟷-𝚏𝚕𝚊𝚜𝚑-𝟷𝟶𝟷𝟽-𝚞𝚕𝚝𝚛𝚊𝚏𝚊𝚜𝚝 for coding, reasoning, and agents.

[Valentin Yeo @valentinyeo]

My personal grok bot clone 😜

Its basically my own harness which uses my regular browser. As I didn't like the idea of having a separate browser only for the harness.

I mirrored a lot of the principles from grok bot as i really liked the reduce first approach. Learned a lot product design wise from grok bot.

Its a mirrored clone of my standard desktop brower and every session has its own window. So chat + tabs stay organized.

It's written in zig. I need to clean it up more but the idea was to super reduce the foot print. @bot is already much better but a lot of the harnesses just take up too ram I find.

I have it on Android too but it uses a different layhout for session management.

[Valentin Yeo @valentinyeo]

@ElectronDev @felixrieseberg @poteto @owengretzinger @o_kwasniewski Yeah agree. I had them not ship to everyone at first so they became another pile up. Now everything within reason goes to everyone right away and I turn stuff off that's bad if it happens.

[Valentin Yeo @valentinyeo]

@blud872 @felixrieseberg @poteto @owengretzinger @o_kwasniewski Exactly 😩 the fix is waiting in the same queue that is blocked 😂

[Valentin Yeo @valentinyeo]

@Castah_ @felixrieseberg @poteto @owengretzinger @o_kwasniewski Felix rieseberg prompt first but the using pstack in combination

[Valentin Yeo @valentinyeo]

@cloneisjun Do you mean the feature flags? Or what do you mean with off switch 🙂

[Valentin Yeo @valentinyeo]

I'm honestly not sure anymore what should stay private on GitHub and what should be open source.

If your bigger project is private, you pay for GitHub Actions minutes. As far as I understand it, that makes CI expensive.

So you either run your own local runners or pay for something like @useblacksmith.

If you ship with agents and push lots of updates every day, that's a lot of minutes. Or a lot of money on outside services.

But so many open source projects became commercial successes.

So early on, I don't think it makes sense to carry CI costs for a project that makes no money yet.