Ever wonder why your coding sessions drag like a bad sequel, while Silicon Valley whispers about devs churning out departments’ worth of work solo?
Boris Cherny, the mind behind Claude Code at Anthropic, just cracked open his terminal playbook. It’s not magic. It’s orchestration — raw, parallel, ruthless.
And here’s the thread that’s lit X ablaze: Cherny’s casual drop of his setup, now dissected by everyone from indie hackers to enterprise leads. Developers aren’t just noting; they’re rebuilding their workflows overnight.
“I run 5 Claudes in parallel in my terminal,” Cherny wrote. “I number my tabs 1-5, and use system notifications to know when a Claude needs input.”
That line alone? Pure dynamite. Forget linear loops — write, test, debug, repeat. Cherny commands a fleet. One agent’s hammering tests. Another’s refactoring cruft from ‘05. A third’s drafting docs that won’t make you weep. iTerm2 pings him like battlefield alerts. smoothly handoffs via ‘teleport’ to browser tabs running 5-10 more Claudes on claude.ai.
It’s Starcraft in code form, as one X user nailed it. You’re not typing syntax; you’re zerg-rushing features.
How Parallel Agents Rewrite Dev’s Inner Loop
Look, traditional coding’s inner loop — that Paul Swail diagram everyone nods at — it’s toast. Cherny shatters it. Five streams? That’s not multitasking; it’s an architectural flip. Humans bottleneck at one brain. AIs? Stack ‘em like servers.
But why five? Not four, not six — five hits the sweet spot. Enough parallelism for real throughput, not so many you drown in context switches. He tabs ‘em 1-5, notifications buzz on output ready. One Claude stalls on a hairy refactor? Pivot to tab 2’s test suite blitz.
This isn’t hype. It’s Anthropic’s “do more with less” gospel, straight from Daniela Amodei. While OpenAI stacks GPUs to the moon, Cherny proves smarts in steering beat brute compute. Exponential gains from orchestration, not just bigger models.
And the kicker — my unique angle here — this mirrors the 1970s terminal revolution. Back then, batch jobs on mainframes gave way to interactive shells. One user, multiple processes, real-time feedback. Cherny’s just ported that to AI: your terminal as the new multi-user system, Claudes as forked processes. History rhymes; dev productivity explodes again.
Short para for punch: Genius.
Why the Slowest Model — Opus 4.5 — Wins Every Sprint
Counterintuitive as hell. Everyone chases latency like it’s oxygen. Faster tokens! Smaller models! Cherny? Nope. Opus 4.5 with ‘thinking’ mode, every damn time.
“It’s the best coding model I’ve ever used, and even though it’s bigger & slower than Sonnet, since you have to steer it less and it’s better at tool use, it is almost always faster than using a smaller model in the end.”
See the why? Bottleneck’s not tokens-per-second. It’s your correction cycles. Dumb model spits garbage — you fix, iterate, rage-quit. Opus thinks deeper upfront, tools like a surgeon. Net speed: your time saved multiplies.
Enterprise suits, listen up. That ‘compute tax’ pays dividends. No more babysitting beta models. Cherny’s data — implicit in his output — screams it: quality cascades to velocity.
But here’s the rub (parenthetical skepticism): Anthropic spins this as accessible wizardry. Truth? It demands Cherny-level finesse. Mere mortals might just burn API credits without the wins.
Three sentences, varied starts. Now sprawl: Imagine scaling this — teams syncing fleets of Claudes, CLAUDE.md as shared canon. Not tomorrow, but soon. Predictions? Solo founders build unicorn MVPs in weeks, not years. Big Tech’s engineer bloat? Obsolete.
The CLAUDE.md Ritual: Codebase That Learns from Screw-Ups
AI amnesia — LLMs forgetting your style session-to-session — kills momentum. Cherny’s fix? One git-tracked file: CLAUDE.md.
“Anytime we see Claude do something incorrectly we add it to the CLAUDE.md, so Claude knows not to do it next time.”
Brilliant. Transforms repo into living memory. Claude ingests it per prompt — no drift. Bad pattern spotted? Append, commit, done. Self-healing codebase.
Why it scales: Git’s versioning bakes in evolution. Team-wide? Everyone’s lessons compound. It’s not prompt engineering; it’s environmental programming. The repo breathes context.
Wander a sec: Reminds me of early version control hacks, RCS files embedding style guides. But AI amps it — dynamic, infinite recall.
Critique time. Corporate spin calls this ‘best practices.’ Nah. It’s battle-tested from Cherny’s trenches. Copy-paste without grokking the rhythm? You’ll flail.
One-word para: Discipline.
Can Regular Devs Pull Off Cherny’s Claude Code Fleet?
Skeptics ask: Is this one-off brilliance, or blueprint for all?
Short answer — yes, with caveats. iTerm2’s free. Claude access? Pro tier. But the mental model shift? Steep. You’re admiral now, not captain.
Tested parallels: Early Vim/Emacs power users felt this — macros as mini-agents. Now AI fills the macros. Devs on X report 3x output already.
Bold call: By Q4, tools emerge auto-fleeting agents. Cursor forks it. Replit integrates. Claude Code workflow goes mainstream, solo devs eclipse teams.
Risk? Over-reliance. AIs hallucinate; fleets amplify. CLAUDE.md mitigates, but humans still gatekeep.
Why This Workflow Signals AI Dev’s Architectural Sea Change
Underlying shift: From code as artifact to process as symphony. Cherny’s not inventing; he’s exposing the OS layer for AI agents. Terminals evolve — notifications, tab orchestration, teleport bridges.
Anthropic wins here. Not model size, but ecosystem glue. OpenAI? Catch up on tooling.
Dense wrap: Productivity’s loop closes — human intent fans to parallel execution, feedback loops tighten via pings and memory files. Result? Software velocity untethered from headcount.
Punchy close: Devs, tab up.
🧬 Related Insights
- Read more: Karpathy’s AutoResearch: AI That Grinds While Humans Snooze
- Read more: AI’s Now Your First Analyst: Welcome to the Uneasy Truce
Frequently Asked Questions
What is Claude Code exactly? Claude Code is Anthropic’s advanced coding agent, powering terminal-based AI workflows for tasks like testing, refactoring, and docs.
How do I set up Boris Cherny’s 5-Claude workflow? Grab iTerm2, spin up 5 tabs numbered 1-5, enable system notifications, and prompt Opus 4.5 instances in parallel — use ‘teleport’ for browser handoffs.
Will Claude Code replace human developers? Not fully — it amplifies them. One dev now matches a team, but orchestration smarts remain human.
