LLenny's Podcast
← All episodes
Boris Cherny19 February 2026

Head of Claude Code: What happens after coding is solved

6Frameworks
15Insights

Frameworks in this episode

Productivity5 steps

Best-Model-First Agent Workflow

Use the most capable model on max effort, start in plan mode, then auto-accept — counterintuitively cheaper.

Read →
Entrepreneurship4 steps

Build for the Model Six Months Out

Design your AI product for the model that ships in six months, not the one you have today.

Read →
Innovation4 steps

Don't Box the Model In

Give the model tools and a goal, not a rigid workflow — scaffolding gains get wiped out by the next model.

Read →
Strategy4 steps

Latent Demand Product Discovery

Find your next product by watching people misuse your current one for something it wasn't designed to do.

Read →
Communication4 steps

The Fast-Feedback Flywheel

Fix every piece of user feedback within minutes so people feel heard and give you even more.

Read →
Leadership4 steps

Under-Resource to Force Automation

Deliberately under-staff projects and hand out unlimited tokens so people are forced to automate with AI.

Read →

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Hot Take· 2

Hot Take17:30

Coding Is Virtually Solved, So Claude Is Starting to Come Up With Ideas

Boris argues that for the kind of programming he does, coding is now virtually a solved problem. The next frontier is Claude generating its own ideas by reading feedback, bug reports, and telemetry, and branching beyond code into general computer tasks like project management and paying a parking ticket.

  • Claude now mines feedback, bug reports and telemetry to propose fixes and things to ship
  • It's becoming less of a tool and more like a co-worker
  • Coding is virtually solved and will become increasingly solved across every stack
  • He uses Cowork daily for non-coding tasks like project management and paying a parking ticket

“Claude is starting to come up with ideas.”

Boris Cherny · 17:30

“at this point it's safe to say that coding is virtually solved.”

Boris Cherny · 18:00
#claude-code#agents#future-of-work#cowork
Hot Take42:30

The Title 'Software Engineer' Will Start to Go Away, Replaced by 'Builder'

Boris predicts the traditional split of engineering, design, and product management will blur, with roughly a 50% overlap already appearing. By the end of the year he expects the title 'software engineer' to start disappearing, replaced by 'builder' or a world where everyone is a product manager and everyone codes.

  • Engineering, design and PM roles show maybe 50% overlap today
  • The three roles persist short-term but are getting murkier
  • 'Software engineer' as a title will start to go away by year's end
  • It may be replaced by 'builder' or by everyone being a PM who codes

“the title software engineer is going to start to go away and it's just going to be replaced by builder”

Boris Cherny · 43:00
#future-of-work#roles#product-management#predictions

Explainer· 3

Explainer05:30

4% of All GitHub Commits Are Now Written by Claude Code

A SemiAnalysis report found 4% of all public GitHub commits are authored by Claude Code, projected to reach a fifth by year's end. Boris says the number understates reality because private repos are higher, and that the truly striking part is the accelerating growth rate, not the current level.

  • SemiAnalysis: 4% of public GitHub commits, predicted to hit ~20% by end of year
  • Private repositories are estimated to be quite a bit higher than 4%
  • Daily active users of Claude Code doubled in just the past month
  • Growth is accelerating across every metric, not just going up

“While we blinked, AI consumed all software development.”

Lenny Rachitsky · 05:30

“4 4% of all commits in the world is just way more than I imagined.”

Boris Cherny · 06:30
#claude-code#growth#software-engineering
Explainer32:30

Why the Printing Press Is the Right Historical Analogue for AI

Boris reaches for the printing press as the closest historical parallel to today's shift. Before Gutenberg, sub-1% of Europeans were literate and scribes did all the reading and writing; afterward printed material exploded, costs collapsed, and literacy eventually spread, democratizing a capability once locked to a tiny elite.

  • In the mid-1400s literacy was under 1%, with scribes doing reading and writing for illiterate lords and kings
  • The 50 years after the printing press produced more printed material than the prior thousand years
  • Printing costs fell roughly 100x over 50 years
  • Literacy rose to about 70% globally over the following 200 years as education systems formed

“the thing that's come closest for me is the printing press.”

Boris Cherny · 32:30

“there was more printed material created than in the in the in the thousand years before.”

Boris Cherny · 33:00
#history#printing-press#democratization#analogy
Explainer47:00

Latent Demand: The Product Principle Behind Marketplace, Dating, and Cowork

Boris explains latent demand, watching people 'abuse' a product to do something it wasn't designed for, as the single most important principle in product. Facebook Marketplace came from 40% of group posts being buying and selling; Facebook Dating from cross-gender profile views; and Cowork from people using Claude Code in a terminal to grow tomatoes, analyze genomes, and recover wedding photos.

  • Latent demand = people hacking a product to do something it wasn't built for
  • Facebook Marketplace: 40% of group posts were buying and selling
  • Facebook Dating: 60% of profile views were opposite-gender non-friends
  • Cowork: people used Claude Code to grow tomatoes, analyze a genome, recover corrupted wedding photos and read an MRI
  • A modern twist: look at what the model is trying to do and make that easier

“latent demand which I think is just the single most important principle in product.”

Boris Cherny · 47:00

“There was someone on Twitter that it to grow tomato plants, there was someone else using it to analyze their genome.”

Boris Cherny · 49:00
#product#latent-demand#cowork#facebook

Story· 5

Story04:00

The Fastest Job Change: Leaving Anthropic for Cursor and Coming Back in 2 Weeks

Boris explains the whiplash job move where he left Anthropic for Cursor and returned two weeks later. He admired the Cursor team and product, but almost immediately realized what he missed was Anthropic's safety-driven mission, which he needs personally to be happy.

  • He joined Cursor because he was a genuine fan of the product and impressed by the team
  • Within days he realized no exciting product could substitute for Anthropic's mission
  • Ask anyone at Anthropic why they're there and the answer is always safety

“It's the fastest job change that I've ever had.”

Boris Cherny · 04:00

“what I really missed about Anthropic was the mission.”

Boris Cherny · 04:30
#career#anthropic#cursor#mission
Story16:00

Shipping 30 PRs a Day: Not a Line Edited by Hand Since November

Boris describes his current workflow: 100% of his code is written by Claude Code, shipping 10 to 30 pull requests every day, without hand-editing a single line since November. He still reviews the code and notes that Claude reviews 100% of Anthropic's pull requests, with a human checkpoint remaining.

  • Ships roughly 10, 20, 30 pull requests every single day
  • Has not edited a line of code by hand since November
  • He still reads the code; correctness and safety require it
  • Claude reviews 100% of pull requests at Anthropic, with a human review layer after

“every day I ship like 10, 20, 30 pull requests, something like that.”

Boris Cherny · 16:00

“I have not edited a single line by hand since uh November.”

Boris Cherny · 16:30
#claude-code#workflow#productivity#code-review
Story22:30

The Memory Leak a Newer Engineer Solved Faster by Just Asking Claude

Boris recounts debugging a Claude Code memory leak the traditional way with heap snapshots and debuggers, while a newer engineer simply asked Claude to figure it out. Claude wrote itself a just-in-time analysis tool, found the issue, and shipped a PR faster than Boris could, illustrating how veterans can get stuck in an outdated mental model of what the model can do.

  • Boris manually took heap snapshots and combed through traces like every engineer has done a thousand times
  • A newer teammate just told Claude 'there's a leak, can you figure it out?'
  • Claude wrote its own just-in-time tool to analyze the snapshot and shipped a PR first
  • Long-time users must transport themselves to the current model; it's not Sonnet 3.5 anymore

“the engineer that was newer on the team just uh had Claude code it.”

Boris Cherny · 23:00

“it found the issue and put up a pull request faster than I could.”

Boris Cherny · 23:30
#debugging#claude-code#mindset#agents
Story21:00

Productivity Per Engineer Up 200% — a Number That's Insane to Anyone in Dev Productivity

Boris contrasts his Meta days, where a year of work by hundreds of engineers on code quality yielded a few percentage points of productivity, with Anthropic today, where per-engineer productivity in pull requests has risen 200%. He stresses how unprecedented and yet how normalized this scale of change has become.

  • Per-engineer productivity in pull requests is up 200% since Claude Code was introduced
  • At Meta, code-quality efforts across hundreds of engineers moved productivity a few percent per year
  • Anthropic roughly 4x'd the engineering team while per-engineer output still jumped
  • The pace of change is easy to normalize but genuinely unprecedented

“productivity per engineer has increased 200% in terms of like pull requests.”

Boris Cherny · 21:00

“you would see a gain of like a few percentage points of productivity, something like this.”

Boris Cherny · 21:30
#productivity#engineering#anthropic#meta
Story1:12:00

Post-AGI, Boris Would Be Making Miso

Asked by Ben Mann what he'd do post-AGI, Boris recalls living in rural Japan as the town's only English speaker, where neighbors bonded by trading pickles and miso. Miso taught him to think on long time scales, months for white miso, years for red, and he says that if he weren't at Anthropic he'd probably be making miso.

  • In rural Japan he was the only engineer and only English speaker in town
  • Neighbors built friendships by trading homemade miso and pickles
  • White miso takes at least 3 months; red miso two to four years, forcing patience
  • The long-time-scale mindset is a deliberate counterpoint to fast-paced engineering

“post-AGI or if I wasn't at Anthropic, I'd probably be making miso.”

Boris Cherny · 1:13:30
#personal#japan#miso#post-agi

Tool· 1

Tool1:08:30

Boris's Claude Code Pro Tips: Use the Best Model and Plan Mode

Boris shares practical Claude Code tips while noting there's no single right way to use it. Use the most capable model (Opus 4.6 with maximum effort) because a cheaper model often burns more tokens overall. Start most tasks in plan mode, which is literally one injected sentence telling the model not to write code yet, then auto-accept edits once the plan is good.

  • Use the most capable model; a less intelligent model often takes more tokens to finish
  • Plan mode is just one sentence injected into the prompt: don't write any code yet
  • In the terminal, plan mode is shift-tab twice; buttons exist in desktop, web and Slack
  • After a good plan, auto-accept edits; with Opus 4.6 it often one-shots correctly
  • Play with different form factors: terminal, desktop, iOS/Android, Slack, it's the same agent

“number one is just use the most capable model.”

Boris Cherny · 1:09:00

“The second one is use plan mode. I start almost all of my tasks in plan mode, maybe like 80%.”

Boris Cherny · 1:09:30
#claude-code#tips#plan-mode#workflow

Takeaway· 4

Takeaway24:00

Underfund Your Teams and Give Engineers as Many Tokens as Possible

Boris shares a counterintuitive management principle: deliberately under-resourcing projects forces people to 'Claudify' their work and ship faster. His advice to CTOs is not to cost-cut early but to hand engineers effectively unlimited tokens so they can try crazy ideas, then optimize only once something works and scales.

  • Under-funding a project forces engineers to automate with Claude to move fast
  • Don't try to optimize or cost-cut at the beginning of an idea
  • Token cost per engineer is usually small relative to salary; optimize only after scale
  • Some Anthropic engineers now spend hundreds of thousands a month in tokens

“when you underfund everything a little bit uh because then people are kind of forced to Claudify.”

Boris Cherny · 24:00

“Start by just giving engineers as many tokens as possible.”

Boris Cherny · 26:00
#management#tokens#productivity#leadership
Takeaway40:30

Be a Generalist: The Engineers Who Cross Disciplines Will Win

Boris's advice for staying ahead is to experiment fearlessly with the tools and, crucially, to become more of a generalist. On the Claude Code team everyone codes, from the PM to the finance person, and the strongest contributors are those who cross product, infrastructure, design, business, and user-facing disciplines.

  • Experiment with the tools, don't be scared, be on the frontier
  • Study more than just CS; the best engineers cross disciplines
  • On the Claude Code team the PM, EM, designer, finance person and data scientist all code
  • Rewarded people will be curious generalists who think about the whole problem, not just the engineering part

“try to be a generalist more than you have in the past.”

Boris Cherny · 41:00

“on the Cloud Code team, everyone codes.”

Boris Cherny · 41:30
#career-advice#generalist#skills#team
Takeaway1:03:30

Don't Box the Model In: Give It Tools and a Goal, Not Rigid Workflows

Boris's advice for building AI products is to stop putting the model in a box with strict step-by-step orchestrators. Give it tools, a goal, and let it figure things out. He invokes Rich Sutton's Bitter Lesson, always bet on the more general model, because scaffolding only buys 10-20% that the next model wipes out.

  • Rigid workflows and orchestrators underperform giving the model tools plus a goal
  • A year ago you needed scaffolding; now you mostly don't
  • Rich Sutton's Bitter Lesson: the more general model always outperforms the more specific one
  • Scaffolding may add 10-20% but those gains get wiped out by the next model, so it's often better to just wait
  • Avoid tiny models and fine-tuning; bet on the general model when you can

“don't try to box the model in.”

Boris Cherny · 1:03:30

“you get better results if you just give the model tools, you give it a goal, and you let it figure it out.”

Boris Cherny · 1:04:00
#ai-products#bitter-lesson#scaffolding#model-design
Takeaway1:05:30

Build for the Model Six Months From Now, Not the Model of Today

Boris says Claude Code's biggest bet in hindsight was building for the model six months out rather than today's. Early on he wrote most of his own code because Sonnet 3.5 wasn't good enough, but the product was designed to shine once the model caught up, which happened with Opus 4 and Sonnet 4. His advice to startups: your PMF will feel bad for six months, then click.

  • From the start, Claude Code bet on the model six months out, not today's
  • Early models like Sonnet 3.5 automated little, so Boris hand-wrote most code
  • The inflection came with Opus 4 and Sonnet 4, when growth went exponential
  • Startups should expect poor product-market fit for the first six months, then hit the ground running
  • Concrete bets: models get better at tools/computers and at running for long unattended periods

“we bet on building for the model 6 months from now.”

Boris Cherny · 1:05:30

“your product market fit won't be very good for the first 6 months.”

Boris Cherny · 1:07:00
#strategy#startups#model-progress#product