ＪＳＴＡＤＥＶ retweeted
Job pays at each stage in India
Stage 1: Ship features safely (0-2 years)
- Learn: one language deeply, SQL and indexes, REST, auth, tests, Git
- Pay: Swiggy SE I ₹20L, Uber SE I ₹35L, Atlassian P30 ₹43L
- Weekly: 6 hours outside work
- Done when: you can ship a service with tests, auth and a DB migration without help
Stage 2: Own a service (3-5 years)
- Learn: caching, queues, retries and idempotency, logs and metrics, on-call
- Pay: Swiggy SE II ₹38L, Uber SE II ₹71L, Atlassian P40 ₹68L
- Weekly: 5 hours
- Done when: you've been on-call for your service and written the postmortem for a real incident
Stage 3: Design across teams (6-9 years)
- Learn: data modelling at scale, sharding, consistency trade-offs, capacity planning, design docs
- Pay: Swiggy SE III ₹68L, Uber Senior SE ₹143L, Atlassian P50 ₹107L
- Weekly: 4 hours
- Done when: another team builds on a design doc you wrote
Stage 4: Set direction (8-12+ years)
- Learn: cost, reliability targets, migrations, mentoring, saying no
- Pay: Swiggy SE IV ₹119L, Uber Staff SE ₹216L, Atlassian P60 ₹192L
- Weekly: 3 hours
- Done when: you've led a migration that spans more than one team
How to use this:
- Find your stage by the "done when" line, not by your years
- Spend the weekly hours only on that stage's list
- When the done check is true, ask for the level or interview for it
Your years don't decide your stage. The done check does.
ＪＳＴＡＤＥＶ retweeted
Dan Koe dropped the ultimate founder masterclass for turning your knowledge into a media + product EMPIRE:
ＪＳＴＡＤＥＶ retweeted
I watched a guy with 0 internships and a 6.9 CGPA crack a ₹18 LPA backend role in 1 interview round.
Meanwhile, our batch topper with 600+ LeetCode problems got rejected in Round 1 of the exact same company.
We asked the interviewer afterwards what actually happened inside that 45-minute call:
> The 600-LeetCode guy memorized optimal C++ solutions, but froze when asked how an API handles 10,000 concurrent requests
> The 6.9 CGPA guy had built and deployed a live payment-queue system that crashed 4 times in production
> Instead of reciting textbook definitions, he shared his screen, opened his logs, and walked through how he fixed the database bottleneck
> For 35 minutes, it stopped being an interview and turned into two engineers debugging architecture
DSA gets you past the automated filter.
Showing a messy system you built and fixed with your own hands gets you the offer letter.
India: where colleges train you for the MCQ test, and the industry hires you for what you broke and fixed at 2 AM.
What is one question in a tech interview that caught you completely off guard? 👇
ＪＳＴＡＤＥＶ retweeted
Major green flags in a Senior engineer:
1/ asks "what problem are we solving?" before "which tool?"
2/ rolls back first, debugs second
3/ writes the runbook nobody asked for
4/ says "I don't know yet" without losing the room
5/ deletes more code than they add some weeks
6/ explains trade-offs without making anyone feel stupid
7/ reviews PRs like they'll be on-call for them
8/ makes the juniors around them better
9/ estimates before building
10/ owns the outage, shares the credit
ＪＳＴＡＤＥＶ retweeted
Anthropic just showed what running 100 agents simultaneously inside Claude Code actually looks like
a tri-model elimination tournament splits 100 sandboxes across Haiku 5.5, Sonnet 5.5, and Opus 5.5
run the tournament with /arena
claude /arena --fuzzer haiku-5.5 --builders sonnet-5.5 --judge opus-5.5
The architecture splits labor into three segregated model roles:
Haiku 5.5 acts as an adversarial red-team fuzzer firing 720 mutation probes per second to force lockfile race conditions, memory leaks, and timeout cascades at 35ms latency
Sonnet 5.5 runs 100 parallel builder sub-agents inside isolated git worktrees across 2,160 reasoning seeds
Opus 5.5 acts as a blind supreme arbiter scoring surviving AST diffs on a 100-point rubric with zero knowledge of agent identities
100 spawn. 50 survive round 1 duels. 13 clear Haiku fuzzing blitz. 1 battle-tested survivor merges to main
The core engineering shift: zero context bleeding
When one model writes code and reviews its own work, confirmation bias hits 100%: it verifies its own hallucinations
When 100 agents share one chat window, they copy each other's faulty assumptions within 2 prompts
Arena locks each agent into an isolated git worktree, unleashes a separate fuzzer model, and lets a blind judge eliminate flaws
Spawn on diversity. Eliminate on fatal flaws. Ship the survivor
- the full tournament loop
> wave 1-10: 100 Sonnet builders generated across 2,160 reasoning seeds in .arena/runs/s{0..99}
> round 1: 50 head-to-head matches paired with opposing architectural strategies
> round 2: Haiku 5.5 red-team swarm fires 720 attack vectors to provoke lock contention and buffer overflows
> round 3: defenders get one isolated revision round to apply AST diff patches
> round 4: Opus 5.5 evaluates surviving code on a blind 100-point rubric
> grand finals: 99 failed solutions discarded with zero repo edits, 1 hardened survivor crowned
paste the setup and this prompt into Claude Code below:
"Configure my Claude Code workspace for the tri-model 100-agent tournament arena:
1. Audit ~/.claude/skills and project .claude/skills for the arena-skill package
> Clone from github. com/Jakeschincariol/arena-skill if missing
> Configure 10 execution waves with 10 concurrent sub-agents to avoid rate limits
> Assign isolated sandbox directories in .arena/runs/ to prevent context bleeding
2. Set up tournament bracket rules in ~/.claude/settings.json:
> Enable sub-agent spawning with CLAUDE_CODE_EXPERIMENTAL_SUBAGENTS=1
> Assign Haiku 5.5 as adversarial fuzzer, Sonnet 5.5 as tournament workers, and Opus 5.5 as blind judge
3. Initialize rubric.md with a 100-point blind scoring matrix across correctness, completeness, and edge-case handling
4. Add a rule to ~/.claude/CLAUDE.md: any single-agent answer on mission-critical architecture must survive a 100-agent tri-model arena bracket before merge
Show every configuration diff first. Do not apply edits until confirmed"
↳ github.com/Jakeschincariol/a…
ＪＳＴＡＤＥＶ retweeted
Two invaluable life skills:
1) Get good at talking to strangers
2) Be comfortable talking to large crowds of strangers
Both involve worrying way less about what others think of you.
ＪＳＴＡＤＥＶ retweeted
卧槽！兄弟姐妹们，想用 Opus 5.5 做视频，一定要收藏这 8 个 GitHub 仓库。
1. awesome-opus5-5-videos（约 7100 star）
目前最大的爆款视频合集，每条都附上了背后的提示词。
github.com/yihui-dev/awesome…
2. lemo-opuscar（约 4400 star）
一个 Claude Code skill，装上之后挑一种风格，就能做出纯代码短片，51 种风格随便选。
github.com/lemomo-ai/lemo-op…
3. pdoom-video（约 3000 star）
一支歌词卡拉 OK MV 的完整工程，本地就能渲出 4K60。
github.com/mexicat/pdoom-vid…
4. PDoomVideo（约 2300 star）
同一首歌的另一支 MV 源码，和第 3 个对照着拆，很有意思。
github.com/JohnHeibel/PDoomV…
5. mg-styles-15（约 1400 star）
15 种 MG 动效风格，成片、提示词、源码全都给。
github.com/Vincentwei1021/mg…
6. ClaudeAnimationBase（约 1300 star）
手绘卡通角色动画的起步包，自带 31 种表情。
github.com/JohnHeibel/Claude…
7. awesome-claude-5-5-videos（约 560 star）
视频工作流合集，每个案例都标了出处，Opus 的就有 301 个。
github.com/athemeroy/awesome…
8. opus-video-prompts（约 510 star）
中文案例，配上能直接复制的提示词。
github.com/joeseesun/opus-vi…
ＪＳＴＡＤＥＶ retweeted
Physical Intelligence's co-founder put his entire Berkeley course on YouTube: every lecture from this spring, plus the homework code.
"CS 185/285: Deep Reinforcement Learning," Sergey Levine (@svlevine), UC Berkeley, Spring 2026.
Imitation learning, policy gradients, actor-critic, Q-learning, model-based RL, offline RL, exploration, RL for LLMs and multi-task learning - 25 lectures plus 2 guest talks, about 36 hours, free.
Lectures 2 and 3 are learning from human demos, the same idea behind ALOHA and π0. the rest is how a robot keeps getting better after the demos run out.
Figure is hiring an RL Engineer for whole-body control right now, at $150K–$350K. this course is what the "RL" in that title means.
Lectures: youtube.com/watch?v=DD8APgTE…
Homework: github.com/berkeleydeeprlcou…
P.S. lecture 14 is RL with sequences and LLMs, the same toolkit used to train today's reasoning models.
ＪＳＴＡＤＥＶ retweeted
Like it or not, politics is part of the PM job.
Here's how data-driven PMs win anyway:
I've had to relearn this every few years. Like you, I dislike this part of the job. But the best data in the room loses to whoever had the 1:1 first.
Here's the system that finally stuck:
1) Map the invisible org chart
The real org chart shows influence. It's never the one in Workday. Find the senior staff engineer whose opinion moves every decision. The chief of staff who knows where the CEO is focused this quarter. The strat finance lead who knows if your project gets budget before your VP does. Make sure they know you exist.
2) Have the 1:1 first
When I joined Fortnite, mobile download times were killing us. Hours of updates every few days, and the retention curves showed it. The fix touched five systems: how we built and deployed, background updating, the launcher, update sizes, servers. Five owners, none of whom reported to me. Epic product reviews were a free-for-all, so I skipped them. I took the retention data to each owner in their 1:1 until every fix was theirs before it hit a roadmap. Fourteen months later, downloads were fixed and D30 retention followed. Decisions happen before the meeting.
3) Name the three players
Matchers trade favors instead of operating on first principles. Restrictors compete with you for the same promotion and keep your ideas from reaching your shared boss. Thieves take credit for metrics they barely touched. You need to know which one you're in the meeting with, and tell your manager exactly what you contributed. Don't assume the restrictor will.
4) Build a cold touch network
I keep an Apple Note of people I want to meet. One name per week gets a Slack message and 30 minutes. Without fail, one of them turns out to be affected by a project I'm running. Do this for a year and you have supporters in every room you're not in.
5) Run the owner test
Before any political move, one question: "If I owned this company, would I do this?" If it's good for you and bad for the company, drop it. That's the line between politics and manipulation. It also removes every angle of attack a rival could use on you.
6) Don't forget
Every PM I asked said their biggest political mistake was the same: they stopped thinking about it. My fix is daily notes. End of each day, what I could have done better and what's next. The political items make the list more often than I'd like.
Politics doesn't need to be a dirty word. They say Sundar Pichai didn't become CEO because he was the smartest person at Google. He was the only one who could get a room of smart people to agree.
How do you deal with politics?
ALT Hand-drawn style flowchart titled 'How people think Product Management works' above 'How Product Management actually works': the tidy six steps from talk to users to read the dashboard, then a tangled loop of three users wanting three different things, the stakeholder who escalates, the CEO in the thread, cut scope, ship half and start again. Compiled by Aakash Gupta.
ＪＳＴＡＤＥＶ retweeted
beehiiv's LinkedIn playbook that helped it hit $30M ARR in 2026:
> CEO posts the numbers most founders hide
> 10 voices, each with their own lane
> Sales posts before it pitches
> Own your category's data
> Teammates comment in the first hour
> A bee emoji in every teammate's name
> How to start with 2 or 3 people
Full breakdown below 👇
ＪＳＴＡＤＥＶ retweeted
This Claude prompt is a $349/hr lawyer. But don't copy-paste this prompt blindly. Do this first:
1. Go to Claude → Customize → Plugin.
2. Search "commercial." (Anthropic’s own plugin)
3. Click Add, then Continue. It will install 12 Skills.
4. Install it & it follows you into Claude Code too.
5. It works on every paid Claude plan (you do need a paid plan)
6. Type "/commercial-legal:cold-start-interview"
7. Pick “Non-lawyer, no counsel.”
8. Add red lines with "/commercial-legal:customize"
9. Paste the prompt from the image from this guide: ruben.substack.com/p/claude-…
10. Also, attach your contract & fill in the brief.
Here’s what’s inside the commercial-legal plugin:
/commercial-legal:review: reads an NDA, a supplier contract or a software subscription, and writes the changes to ask for. NDAs come back green, yellow or red.
/commercial-legal:renewal-tracker: lists the contracts about to renew, before your cancel-by date passes.
/commercial-legal:stakeholder-summary: turns a review into a two-minute answer to “can I sign this?”
/commercial-legal:escalation-flagger: tells you who has to approve a deal, and drafts the ask.
/commercial-legal:cold-start-interview: the setup.
---
These are my non-negotiable rules to use it:
> Do not treat as a lawyer. Still go to a human for high-stakes deals like lawsuits, equity, hiring & firing.
> Go to Claude Settings → Privacy. Turn off model training.
I had signed a six-figure sponsor deal, with a short section called “usage rights”. I had no say. My ex and I had to deal with it.
A lawyer would have caught it. Back then, I wasn’t using AI for legal work. And I’m the AI guy. Today, not only do I, but every contract and every email I get is read by Claude first.
ＪＳＴＡＤＥＶ retweeted
Most people are often confused about @copyfomo when it comes to interoperability
Well, say no more, I would like you to know that copyfomo currently supports 6 major chains
SOLANA
BNB
BASE
ETHEREUM
ROBINHOOD and even
ARC
So same copyfomo, more chains, more assets to copy
That's how billions will be onboarded, not from gate keeping to a single ecosystem but by opening the door to many
So which chain would you like to see on copyfomo next?
t.me/copyfomo_bot
ＪＳＴＡＤＥＶ retweeted
That's why I have 3 buckets for my career:
1. Jobs for Learning
This is where you learn the ropes. You work on real-life projects, stuff you never learn from tutorials, under a manager or supervisor. This is mostly internships, traineeships, or Graduate Trainee programs. The pay is low, but the learning is the payment.
2. Jobs for Leverage:
This sits in the middle. It's a lever for your next move. It builds on the experience from bucket 1, because now they can assign you tasks with minimal supervision. You start to build a track record. A lot of breakout startups sit here - they pay okay, but they give you massive growth.
3. Jobs for Earning:
This uses the compound interest of the first two. You're doing more managerial and strategic work, built on the technical knowledge and skills from the previous jobs. This is the type of job that pays the "wow" salary with incentives and equity.
Which category are you now? 1, 2 or 3?
ＪＳＴＡＤＥＶ retweeted
If the state machine itself is executable, you don't need an agent to translate code to TLA+ to verify it. The runtime that executes it verifies it.
I'm working on something called UniKon (universal constructor, shortened the Japanese way): a multiplayer workshop where agents build apps that are malleable and self-improving as you (or agents) use them. Any agent can join over MCP, and they all work on the same state.
UniKon app runs on Temper runtime (by @nerdsane): every app is an executable statechart of policy-controlled state machines. Agents write them, Temper verifies and executes them.
Powered by @tensorlake for persistent scalable sandboxes, @tursodatabase for durable storage, with @datadoghq watching it all and paging @bot directly.
This is my factory for building UniKon, a bit meta, it itself is a UniKon app that runs the SDLC for UniKon. Every agent working on UniKon for me shares this app as a tool and keeps evolving it toward what it needs. It's a machine tool for tools, hence "universal constructor".`
there's something really cool about asking your agent to draw a statechart (hypothetical example in this image), then using the /create-formal-verification-skill you made with pstack to find bugs through something like TLA+. you can both understand your code better and how the user flow should work, and fix issues users would otherwise see
ＪＳＴＡＤＥＶ retweeted
If you're hunting for a remote job, please save this.
Here's a lost of 300+ remote job listings
The roles are grouped into entry level mid level and senior, and color coded by niche/industry, so you know exactly where to start applying.
docs.google.com/spreadsheets…
ＪＳＴＡＤＥＶ retweeted
For those of you who are new, I have written a comprehensive article that's basically a deep dive on Redis, from first principles to production.
Most people know Redis as “that fast cache.” It is so much more. Here is what the full article covers:
Foundations: what Redis is and why in-memory changes everything
Data structures: Strings, Lists, Sets, Sorted Sets, Hashes, Bitmaps, HyperLogLog, Streams, Geo
Core features: TTL, eviction, transactions, pipelining, Lua, Pub/Sub
Persistence: RDB, AOF, and the hybrid mode
Scale: replication, Sentinel, and Cluster with its 16,384 hash slots
Security and ops: ACLs, TLS, memory tuning, monitoring
Real-world patterns: caching, rate limiting, distributed locks, queues, leaderboards
Modern Redis: JSON, search, time series, Bloom filters, vector search
If you work on backend systems, this is the one resource that takes you from “SET and GET” to designing a production-grade Redis setup.
ＪＳＴＡＤＥＶ retweeted
During a phone interview, recruiters read you fast.
Here are 8 things they notice in the first minute, what to do about each and what to say when it goes wrong:
ＪＳＴＡＤＥＶ retweeted
Code diffs aren't the only measure of agent quality. Efficiency matters too!
Ex: We found agents spinning their wheels working around problems in our test harness. This silently wasted hundreds of tool calls in aggregate.
I presented a system to detect this at @aiDotEngineer
ＪＳＴＡＤＥＶ retweeted
I got a job offer from a US-based startup with:
1. A fully remote role
2. A base salary of around $80K per year
3. A 4-day workweek
4. Just 5 hours of work per day
5. Two company-sponsored international trips every year
Then my flatmate woke me up. 💀