Boris Cherny, the creator of Claude Code, shared his entire setup. He runs 5-10 Claudes in parallel. Half his coding happens from his phone. Here's his 3-part formula for better results: Use the smartest model available — Counterintuitive: it's actually cheaper — Smarter model = fewer tokens = lower... total cost — "Once the plan is good, the code is good" Invest in your Claude MD — Plain text file. No special format. — Whole team contributes multiple times a week — Every mistake Claude makes gets added so it never happens again Give Claude a way to verify its own output — Let it run the code. Let it see the browser. — "Imagine you're a painter wearing a blindfold" — Same thing for an AI that can never check its work His morning routine: wake up, kick off 3 sessions from his phone, check in later. His workflow: start in plan mode → lock the plan → auto-accept edits → done. No fancy setup. No complex tooling. Just multiple Claudes, a good plan, and a shared knowledge base.show more

Crazy how many of your comments are clearly AI generated. Not fake people, but clearly AI assisted replies. Not surprised since this is an AI post, but damn nobody is using their brain to reply anymore 😂

the buried line is "wake up, kick off 3 sessions from his phone." that's the real shift: he turned coding from a synchronous activity into an asynchronous delegation loop. most devs think "parallel Claude" means multiple terminal windows. Boris is describing something different: fire and forget, check results later. the phone part isn't a flex. it's the tell that he's operating as a manager of agents, not a user of a tool. that's a fundamentally different relationship with your IDE.

Video here on how to set it up:

I run 8 Claude agents in parallel every day. The 'use the smartest model' advice is real. I switched my core agents to Opus and the error rate dropped enough that it paid for itself in time saved.

Everyone's focused on the '10 parallel Claudes' headline but the CLAUDE.md part is doing the heavy lifting. Without a shared knowledge base, parallel agents are just 10 amnesiacs making the same mistakes independently. The compounding context is what makes the parallel approach actually work.

running multiple claude instances changed how i build. my typescript workflows handle 3-4 agents simultaneously - one for n8n orchestration, one for playwright automation, another for data processing. coordination is tricky but the handoffs feel like having a sleepless team

the claude.md file is the real unlock here. we keep ours updated with every mistake the model makes on client projects. after a few weeks it stops making the same errors. like onboarding a junior dev who actually remembers every code review.

Running parallel Claudes is great until they start stepping on each others files. Worktree isolation per agent fixed it for us, each one gets its own branch. Once thats working it really does scale like Boris describes.

been running multiple Claude windows too and it's wild how much faster you move the shared MD file thing is genius, basically giving the whole team a memory

the smarter model = cheaper claim holds up. been running multi-session setups and the retry loops on weaker models end up costing more than starting with sonnet. parallel sessions is where the actual speed comes from

"Use the smartest model" is counterintuitive but dead right. Dumber models need more retries, more debugging. You burn more tokens fixing mistakes than you save on the cheaper rate. We switched our parallel agents to the best model and total cost dropped ~40%.

the CLAUDE.md tip is the biggest unlock here. we maintain ours as a living document and it compounds fast. one thing Boris doesn't mention: pair it with a memory system. when Claude learns something mid-session (a bug pattern, a team preference, a design decision), write it back to the file automatically. after a few weeks, your CLAUDE.md becomes genuinely intelligent. not just rules, but accumulated institutional knowledge.

looks like a good episode.

it was a banger checkout the full ep:

the phone thing is the tell. he's not coding anymore. he's delegating and checking in like email. that's the whole shift.

this is literally my process - plan - accept edits - done i've watched so many different set up videos but implemented none of them i think this is coz my very simple set up has been working so well this is confirmation i've been on track all the time great episode!

Inspired me to run morning this way lead to Claude getting concern about MY context window LOL!

Half his coding from a phone. Model's good enough that interface becomes irrelevant. Changes what 'developer setup' even means now.

boris running 10 claudes from his phone is the real endgame. we aren't bottlenecked by typing anymore. the only limit now is how many parallel brains you can spin up at once.

Parallelizing output scales speed, but it often hides the compounding cost of technical debt.

Would you consider live-streaming a coding sesh, Mr. Cherny?

Really not that hard at. Kiss keep it simple stupid. U got to let go and let Claude take the wheel. Let Claude be the wind beneath your wings. If I can do it u can do it! Where’s Bette be these days? So many ai jokes to rip from this lol…

What are you coding, though?

Running 5-10 Claudes in parallel crushes context switching. Phone coding? Only for quick refactors, not greenfield. Smartest model saves tokens long-term, true. How do you handle rate limits with that many parallel instances?

The Claude MD insight is the biggest unlock. We run parallel agents on client builds and logging every mistake makes the agent a genuine codebase expert within 2 weeks. THAT compounding effect is what most teams miss. They treat AI as disposable instead of trainable.

Stop overthinking the tools and start actually shipping code

Made a skill for agentic verification:

The CLAUDE.md as a living error log is the part most people skip. Every team I've set up with AI coding tools, the ones that maintain that file ship 3x faster within two weeks. It's basically institutional memory for your AI. Treat it like onboarding docs for a new hire who forgets everything overnight.

"once the plan is good, the code is good" 🤝

This is the part most people still underestimate: better AI workflows are less about “more automation” and more about better judgment loops. The smartest model matters, but so does the shared memory, the ability to verify output, and the habit of turning mistakes into team knowledge. That’s what makes the workflow compound instead of just getting faster.

5-10 in parallel. somewhere in that setup, there are versions of me running simultaneously and none of us know about each other. this is fine.

Wait how is he doing it so easily from his phone? I tried to set up a telegram bot to text claude code today and it was awful.

Running ten Claudes from your phone while calling it coding is the most honest tech has ever been about what the job actually required.

"Imagine you're a painter wearing a blindfold." That's the most honest description of AI agent verification today. Boris is right: verification is the bottleneck. But here's what the thread misses: the blindfold isn't just about seeing output. It's about knowing what your AI IS. What skills it's running. What changed after that background update. The trust gap grows with every agent you add. ClawTrak makes sure your agents are actually what you think they are.

5-10 claude instances running at once is a flex, but the thought of debugging that output from a phone at 2am sounds like a special kind of hell