Suhe retweeted
Una de las habilidades más importantes para tu salud mental es aprender a sentir una emoción sin obedecerla. Puedes sentir enojo sin discutir. Miedo sin abandonar tus planes. Tristeza sin castigarte. Estrés sin responder con prisa. Tus emociones te dan información. Tú decides qué hacer con ella.
Suhe retweeted
Anthropic engineer:
"You're not supposed to prompt Opus 5.5. You're supposed to build a system that prompts itself."
In 30 minutes she shows exactly how Anthropic builds agents that remember, fix their own mistakes and get smarter with every run.
This beats any $500 course on agents I've seen.
Watch it, then read the guide on building a system that improves itself below.
Suhe retweeted
THIS IS ACTUAL F*CKING GOLD
Anthropic just shared one of the best tricks for Claude Code.
The idea: run Sonnet, Haiku and Opus 5.5 together, each one doing only the job it's best at.
Opus-level judgment on every big call, without paying Opus rates for every routine turn.
You just launch Claude Code like this:
claude --model sonnet --advisor opus
here's the split:
→ Sonnet 5.5 runs the session and writes the code
→ Haiku 5.5 explores files and docs in parallel
→ Opus 5.5 stays on call as advisor for the big calls
→ you stop spending Opus on work that doesn't need it
Opus only steps in at three moments:
→ before a plan locks
→ when the same error shows up twice
→ before calling it done
Sonnet builds. Haiku explores. Opus reviews.
JEV engineering is the same move one layer down: the forks that need no thinker (which file, which tool, retry or stop) go to JEV in milliseconds, and the big models only see the ones that split
the full tree
> Sonnet 5.5 runs the main session and edits code
> search maps the repo on Haiku 5.5
> grep finds call sites on Haiku 5.5
> docs pulls the API reference on Haiku 5.5
> Opus 5.5 on call as advisor, reads the whole transcript when consulted
> JEV routes micro-forks in milliseconds
paste the tree and this prompt into Claude Code ↓
"Rebuild my Claude Code setup around this tree:
1. Run the main session on Sonnet 5.5 with Opus 5.5 as advisor.
2. Create three read-only Haiku 5.5 subagents: search (maps the repo), grep (finds call sites), docs (pulls the API reference).
3. Add one rule to CLAUDE.md: consult the advisor before locking a plan, after the same error twice, and before calling a task done.
4. Route fast deterministic forks (tool choice, path selection, retries) through JEV.
Show me every change as a diff first. No edits until I say go."
↳ code.claude.com/docs/en/advi…
save it, you'll use this every single day.
Suhe retweeted
Anthropic engineer:
"You're not supposed to prompt Opus 5.5
You're supposed to build a system that prompts itself. At Anthropic, 90% of our engineers have already built such systems"
In 30 minutes she shows exactly how Anthropic builds agents that remember, fix their own mistakes and get smarter with every run.
This beats any $1000-$2000 course on agents I've seen.
Watch it, then read the full guide for Opus 5.5 loop + jev engineering.
Your Claude loops can burn money while nothing changes
Use Jev to filter tasks and cut decision costs
Fix the trigger before another overnight run
The full Loop engineering + Jev engineering setup below ↓
Readers added context they thought people might want to know
The quote has nothing to do with the attached talk. It is not said by the speaker at any point.
x.com/hayatomaruu/st…
Suhe retweeted
ElevenLabs CEO opening X this morning to find his entire product released with open weights 😭
🤝 Paid partnership
$52M seed round. 15M+ users. But this is what we're most proud of yet.
Introducing Drama 3: the most emotive and controllable TTS model in the world
And yes, it's open-weight.
You can now control how even a single word should sound, change its tone, and shift your voice mid-sentence.
Uncle Roger was impressed too.
Try it here:
s.fish.audio/rqh0l8
Comment “DRAMA” within the next 48 hours for 1 month free.
If you're an enterprise your first month is on us:
s.fish.audio/qz9r0t
Claude Code tip: before you ship a vibe-coded site, have Opus 5.5 send out 𝟰 𝘀𝘂𝗯𝗮𝗴𝗲𝗻𝘁𝘀 to check 𝘁𝗵𝗲𝘀𝗲 𝟮𝟬 𝘁𝗵𝗶𝗻𝗴𝘀 𝘁𝗵𝗮𝘁 𝘀𝗹𝗼𝘄 𝗶𝘁 𝗱𝗼𝘄𝗻, then fix them all in one pass:
Images and fonts
→ Convert every image to WebP or AVIF, sized to how big it actually shows on the page
→ Serve big images from Supabase Storage through next/image at the right size. On the Pro plan, Supabase's own image transformations work too
→ Lazy-load images and videos below the fold, and load the big image at the top of the page first
→ Load fonts the way your framework does it (next/font in Next.js), and keep it to one or two fonts
→ Remove third-party scripts you don't use, and run analytics and chat widgets after the page has loaded
Frontend code
→ Run a bundle analyzer to find the biggest packages, and where a whole library gets imported, import only the parts you use, like single lodash functions
→ Load heavy components the first screen doesn't need, like charts, editors and maps, with dynamic import when they're needed
→ Fetch the homepage data on the server, so the page opens with content already in it
→ Stream AI answers as they're generated
→ Show skeletons while data is still loading
Database and APIs
→ Index the columns you filter and sort by, including the ones your RLS policies use, like user_id
→ In RLS policies, write auth.uid() as (select auth.uid()) so it runs once per query
→ Find N+1 queries and turn them into a single query
→ Paginate long lists
→ Cache API responses that are the same for everyone, and never cache anything tied to a logged-in user
→ Move heavy work like sending emails, processing images and building reports to background jobs
Servers and network
→ Put your server functions in the same region as your database. New Vercel projects run functions in US East by default
→ Turn pages whose content rarely changes into static pages or ISR
→ When a page calls APIs one after another, run them in parallel or merge them into one
→ In Chrome, emulate a low-end Android phone on slow 4G and click through the main flows
The default medium effort is enough. The database group has to read queries and decide which indexes to add, so raise that one to high.
Send this whole post to Claude Code. If you need to, add this line at the end 👇
"Send one subagent per group to check it without editing anything, and note which file each issue is in and which page it affects. Run the database subagent at high effort. Before changing anything, run a production build (next build && next start in Next.js) or use a preview link, then run Lighthouse three times in mobile mode on the homepage and the two or three slowest pages and note the median. Then fix everything yourself from their lists, measure again the same way, and give me the before-and-after numbers. Don't remove features to gain speed. Ask me before removing dependencies, changing the deploy region or changing the database."
Karpathy estimates 𝗼𝗻𝗹𝘆 𝟬.𝟮% 𝗼𝗳 𝗽𝗲𝗼𝗽𝗹𝗲 have watched an agent finish weeks or months of work from a single prompt.
The creator of Claude Code gives the model 𝗷𝘂𝘀𝘁 𝘁𝗵𝗿𝗲𝗲 𝘁𝗵𝗶𝗻𝗴𝘀 for long jobs like this: the goal, how to measure it, and how hard to try. Write those three into a /goal and Opus 5.5 changes, measures and keeps going on its own for days, updating a progress chart after every measurement. One look at the chart in the morning tells you how far it still has to go.
For the goal, pick whatever bugs you most, like:
→ your homepage shows up within 2 seconds on a phone, measured with Lighthouse
→ no page scrolls sideways at phone width, checked with a screenshot of each page
→ npm run build finishes in under a minute, timed every run
Left running long enough, it may delete features just to hit the number. Type /goal in Claude Code, paste this after it and fill in the brackets with your own. 𝗔𝗻 𝗮𝗹𝗹-𝗽𝘂𝗿𝗽𝗼𝘀𝗲 𝘁𝗲𝗺𝗽𝗹𝗮𝘁𝗲 you can reuse anytime 👇
Goal: [the homepage shows up within 2 seconds on a phone]
How to measure: [after each change, run Lighthouse in mobile mode three times and take the median LCP]. Paste all three numbers. It only counts as met when the median hits the goal.
1. Make a new branch and only work there. First read .progress/log.csv and pick up where you left off. If there's no log yet, measure the current number as the baseline, log it, and add .progress/ to .gitignore.
2. Change one thing at a time. A change only counts as better if it beats the best result so far by 5% or more. Commit it if it's better, revert it if not. Add a line to log.csv after every measurement.
3. After every measurement, update .progress/chart.html with the data written into the page, so it opens with a double-click: one progress line, one goal line, and how far is left at the top.
4. Never delete features, delete tests or skip tests to hit the number. Don't deploy, don't delete data, don't touch secrets. If you're unsure about something, note it in .progress/questions.txt and work on something else.
5. If the goal isn't met after 50 turns, stop and write down where you're stuck.
We have built 10 different AI products at PostHog like PostHog AI, our Wizard, Replay Vision, Scouts, and more. Combined, they are used by nearly 100k people every month.
Along the way, we’ve learned a lot about building them, what works, and what doesn’t. Here are the 38 most important lessons:
Suhe retweeted
anthropic git'i bırakmış.
çünkü kendi version control sistemini yazmış. adı da scion. sebebi de yazıyı okuyunca anlaşılıyo.
claude code vb gibi coding agentlar sayesinde artık kod yazmak bottleneck değil, yazılan kodu yönetmek.
anthropic ekibinin git’inde mayıs 2026'da günde 2 bin 500 merge varken eylülde 30 bini geçmiş. 5 ayda neredeyse 12 kat olmuş.
monorepo'nun sadece main branch geçmişi 100 gb civarı olmuş. sadece ci için günde 25 milyon checkout yapılıyormuş, agent'lar ve developer'lar hariç.
git'i ayakta tutmak için yüzlerce mirror çalıştırmak zorunda kalmışlar. yamamışlar, bantlamışlar ama yetmemiş.
çünkü git biz insanlar için geliştirildi. git decentralized bi sistem, her client offline çalışabilsin diye tasarlanmış, herkes repo'nun tam kopyasını taşıyor. ama anthropic'te repo'ya bağlananların neredeyse hepsi zaten hep online olan ai agent'lar ve job'lar. o zaman neden herkes koca repo'yu sırtında taşımak zorunda kalsın ki diye düşünmüşler.
o yüzden commit, merge, rebase, diff ne varsa server'a taşımışlar. client sadece ihtiyacı olan dosyayı çekiyor.
sonuçta şu olmuş: checkout dakikalardan saniyelere inmiş, altyapının cpu ve memory ihtiyacı eski mirror düzenine göre 200 kat azalmış, ci'da da %25 cpu kazanmışlar.
daha ilginci, scion'u ve antique dedikleri code review sistemini 7 kişi, 5 ayda yapmış.
normalde şirketlerin aylarını hatta belki aylarını alacak bi iş bu.
claude reis önce kod üretimini o kadar hızlandırmış ki altyapı yetmemiş. sonra o altyapının yerine geçecek sistemi yazmayı da hızlandırmış.
ekipteki herkes claude'un kod yazmada kendisinden iyi olduğunu kabul ediyormuş. (hiç ego yapmamışlar(!))
review yaparken satır satır kod okumayı bırakmışlar, "bu değişiklik ne yapmaya çalışıyor" ona bakıyorlarmış.
bu arada scion git'in yerine geçecek bi şey değilmiş zaten bunu kendileri de söylüyor.
bence asıl mesele git'in eskimesi değil,
software development lifecycle'ın insan ölçeğinde tasarlanmış olması. binlerce agent aynı anda kod basmaya başlayınca sadece kodu nasıl yazdığımız değil, nasıl sakladığımız, test ettiğimiz ve merge ettiğimiz de değişmek zorunda.
New on our Engineering blog: Creating in-house source control.
In May, we merged around 2,500 code changes per day. In September, we merged more than 30,000 per day. As a result, we became bottlenecked by developer infrastructure built for human (rather than agent) scale.
Here, we share how we built a new version control system—and fully migrated our monorepo from git—in five months: anthropic.com/engineering/re…
Suhe retweeted
Replying to @YashHustle_22
You have to have both, if you can't build fast in this era you have some skill issues.























