Create a detailed outline of this interview. Then, construct a  timeline of events in the development of Claude Code. Then, create a biography of the guest. Lastly, make a list of memorable quotes from all the main points and relevant moments in the interview.

1. Detailed outline of the interview

0:00–3:45 – Cold open and setup

Boris states that 100% of his code has been written by Claude Code since November, shipping 10–30 PRs per day, with multiple agents running in parallel while he works.
He says he enjoys coding more now because he avoids minutiae, and that productivity per engineer at Anthropic is up roughly 200%.youtube+1
Boris argues coding is “largely solved” and imagines a world where everyone can program and the “software engineer” title gives way to “builder,” with everyone effectively product managing and coding.youtube+1
Lenny frames Claude Code as nearing its one‑year anniversary, notes it has transformed software engineering and is now reshaping other functions, and mentions Anthropic’s rapid growth and funding.youtube+1

3:45–8:41 – Leaving Anthropic for Cursor and returning

Lenny asks about Boris’s brief move to Cursor and return to Anthropic after two weeks.
Boris: joined Cursor because he admired the team and product and thought they saw AI coding’s trajectory early.
Realized quickly he missed Anthropic’s mission: building safe AI. Everyone at Anthropic, if asked why they’re there, answers “safety,” and he needs that mission-driven context to be happy.youtube+1
Transition into reflecting on one year of Claude Code and its impact.

5:35–13:29 – One year of Claude Code: growth, origin story

Lenny cites SemiAnalysis reporting that Claude Code now authors ~4% of all public GitHub commits and could reach ~20% by year end, and mentions a Spotify report that its best developers haven’t written code since December because of AI.
Boris finds the 4% figure “totally crazy,” notes the real number is higher including private repos, and emphasizes the acceleration of growth (DAUs doubling in the past month).
Origin story:
- Anthropic labs team strategy: first models great at coding, then tool use, then computer use, in service of safety.
- Boris joined Anthropic, spent a month hacking prototypes, another month doing post‑training research “to understand the layer under the layer,” then returned to prototyping what became Claude Code.youtube+1
- First prototype “ClaudeCLI” was a terminal tool using a few tools; shock moment: with only a generic “bash” tool, the model figured out how to tell him what music he was listening to, purely by tool use.
- He posted an internal demo that got only two likes; people thought real coding tools had to be IDE‑based, not terminal‑based.
Product lesson: under‑resource early; he kept it terminal‑based because he was alone and needed a fast, flexible form factor while the model was improving too quickly for heavy UI.youtube+1

13:29–22:24 – Exponential change in engineering and agents

Claude Code slowly became internally popular; DAUs went vertical after Ben Mann pushed Boris to track DAUs early.
External launch: not an immediate hit; early adopters loved it but broader understanding took months because it was so different and terminal‑based.
Now Claude Code is everywhere: terminal, mobile, desktop, web, IDE extensions, Slack, GitHub.
Reflection: huge humility and learning from users; no one really knows what they’re doing, and user feedback is the main signal.
Lenny highlights how mainstream AI coding has suddenly become; predictions about “100% AI‑written code” that once seemed crazy now feel obvious.
Boris recounts a May 2025 prediction at Anthropic’s developer conference (“by end of year you might not need an IDE”), which drew gasps, but followed from Anthropic’s exponential/s scaling-laws mindset and extrapolating % code written by Claude.youtube+1
By November, Claude was writing 100% of Boris’s code; now many customers see similar patterns.

22:24–32:15 – Experimentation, innovation, and 100% AI‑written code

Discussion of experimentation: innovation can’t be forced; you must give people space, psychological safety to fail, and accountability to kill bad ideas quickly.
Early Claude Code wrote a small share of Boris’s code (20% in February, 30% in May); it crossed 100% only in November, but from day one, he felt he was “onto something” and spent nights and weekends hacking.
Current workflow: 100% of his code is written by Claude Code; he still reads code and relies on Claude’s automatic code review plus human review, especially for production code.youtube+1
Next frontier: Claude now:
- Mines feedback, bug reports, telemetry, and proposes ideas for bug fixes and features — acting more like a co‑worker.
- Extends beyond coding into general tasks via Cowork agents, automating things like paying parking tickets and managing team project logistics across spreadsheets, Slack, and email.
Boris asserts coding is “largely solved” for his type of programming and expects this to become true across stacks in coming months; the new frontier is deciding what to build.youtube+1

32:15–44:37 – Should people still learn to code? Printing press analogy

Lenny asks if people should still learn to code.
Boris:
- Today, users of agents still need to understand the layer underneath, but in 1–2 years that likely won’t matter.
- Historical analogy: printing press and literacy. Before Gutenberg, <1% could read/write; within ~200 years of mass printing, literacy rose toward ~70%. The printing press decreased cost of printed material ~100x in 50 years and enabled the Renaissance by democratizing knowledge.youtube+1
- Scribes initially welcomed the press because it freed them from tedious copying, letting them focus on art and binding; he feels similarly about AI coding freeing him from tedious implementation.
He envisions a future where everyone can program; we can’t foresee what that unlocks, just as 1400s Europeans couldn’t foresee modern society.
Acknowledges disruption and pain for many; stresses need for a societal conversation and Anthropic’s work with economists, policy, and social-impact staff.youtube+1

44:37–54:04 – Advice for individuals in the AI era; roles beyond engineering

Lenny reports informal polls: ~70% of engineers and PMs enjoy work more with AI; only ~55% of designers do, with a larger minority enjoying work less.
Boris notes Anthropic hires technically strong people even for non‑technical roles; many designers code and use Claude Code and the desktop app heavily, allowing them to unblock themselves without engineers.
Advice to individuals:
- Immerse yourself in AI tools, be on the frontier, don’t fear them.
- Become more of a generalist: on the Claude Code team, everyone codes (PM, EM, designer, finance, data science).
- Most effective people combine disciplines (e.g., product+infra, product+design, engineering+business, engineering+user research).
On the future of roles:
- Today, PM/design/engineering still exist, but there’s ~50% overlap, and many people do similar work with different emphases.
- By end of the year, he expects titles to blur further; “software engineer” may be replaced by “builder,” and everyone will code and do product.

54:04–59:35 – Latent demand; Cowork built in 10 days

Boris explains “latent demand” in product development: users misuse or “abuse” a product for something it wasn’t designed for, revealing a strong unmet need.
Examples from Facebook:
- Marketplace emerged from observing that ~40% of posts in Facebook groups were buy/sell.
- Facebook Dating emerged from seeing that ~60% of profile views were non‑friends of opposite genders.
Latent demand at Anthropic:
- Non‑engineers (data scientists, etc.) used terminal‑based Claude Code for tasks like SQL analysis, genomics, MRI analysis, gardening, and disk recovery — jumping through hoops to use a terminal.
- This indicated demand for a non‑technical, task‑oriented agent product, leading to Cowork.
“Latent demand from the model”: observe what the model tries to do when given tools and minimal scaffolding; build product around that instead of boxing it into narrow workflows.
Cowork build story:
- Team explored options for months; final idea was “take Claude Code and put it in the desktop app,” with strong security and guardrails.
- Cowork was built in ~10 days, with Claude Code writing essentially all of the implementation, including a full VM and security scaffolding; launched early as a research preview to learn from users and safety behavior.

59:35–1:14:02 – AI safety, mechanistic interpretability, race to the top

Anthropic’s three-layer safety stack:
1. Alignment & mechanistic interpretability: study internal neurons, detect concepts like deception, and monitor activations.
2. Evals: “petri dish” lab tests of the model in synthetic scenarios.
3. In-the-wild behavior: observing how products like Claude Code and Cowork behave with real users.
Claude Code ran internally for 4–5 months before launch to study agent safety; lessons fed back into model and product.
Cowork follows a similar path: aligned and good on evals and internal testing, but needs real-world observation; hence “research preview.”
Boris references Chris Olah and mechanistic interpretability work; notes phenomena like superposition, where concepts are encoded across neurons.youtube+1
Anthropic pursues a “race to the top”: open-sourcing safety infrastructure like a sandbox that any agent can run in, encouraging others to adopt similar safeguards.

59:35–1:03:21 – Agent anxiety, work patterns, personal background

Lenny raises “agent anxiety”: people feel uneasy when their agents aren’t running, fearing lost productivity.
Boris typically has multiple agents running across terminal, desktop, and iOS; he uses mobile heavily for coding, which surprises him.youtube+1
He still regards this as “coding” — describing what he wants instead of writing syntax — likening it to prior transitions (punch cards → software).youtube+1
Shares family background: born in Odessa, Ukraine; his grandfather was among the first programmers in the Soviet Union, working with punch cards.
Lenny reveals he was also born in Odessa; they trade brief emigration stories and gratitude for growing up in the US.youtube+1

1:03:21–1:12:13 – Building AI products: tokens, not boxing the model, bitter lesson

Advice for building AI products:
- Underfund teams slightly to force automation (“claudify everything”); one engineer with agents can ship extremely quickly.youtube+1
- Encourage speed: if something can ship today, do it today; using Claude more is the best way to move faster.
- Give engineers lots of tokens rather than optimizing costs early; early exploration is cheap relative to salaries, and you only optimize token costs once an idea proves valuable.youtube+1
Principle: don’t over‑orchestrate or box the model in with rigid workflows; instead, give it tools and a goal, let it decide steps and retrieve needed context.youtube+1
“Bitter lesson” (Rich Sutton): general models outperform highly specialized ones over time; corollary: bet on the most general, capable model instead of micro‑optimizing tiny/fine‑tuned models and heavy scaffolding, which tends to be obsolete with each new model generation.youtube+1
Build for the model six months in the future: early Claude Code didn’t write much code, but the product was designed for the day when the model could do most of the work; when Opus 4 and Sonnet 4 arrived, usage inflected hard.youtube+1
For external builders: infer likely trajectories (better tool use, longer unattended runs) and design products assuming those capabilities will soon exist.youtube+1

1:12:13–1:21:00 – Pro tips for using Claude Code, Codex comments, post‑AGI plans

Practical tips for Claude Code users:
1. Use the most capable model (currently Opus 4.6) and “maximum effort,” as it often uses fewer tokens overall by needing less correction.
2. Use “plan mode” for most tasks: insert a single instruction (“please don’t write code yet”), iterate on the plan, then auto‑accept edits to one‑shot the implementation.
3. Explore different interfaces (terminal, desktop app, mobile, Slack, web), find what fits your workflow; it’s the same agent everywhere.
On competing coding tools like Codex: Boris hasn’t used it much, but sees competition as healthy; the Claude team focuses on user problems rather than competitor products.
Asked about “post‑AGI” life, Boris describes a previous life in rural Japan, making miso and living slowly; post‑AGI he imagines returning to that kind of long‑timescale craft, focusing on miso making.youtube+1

1:21:00–1:27:45 – Closing reflections and lightning round

Boris reiterates Anthropic’s long‑standing strategy: start from coding, then tool use, then full computer use, both for capability and safety learning.
Claude Code has become a multi‑billion dollar business and major driver of Anthropic revenue; yet he still sees the current state as “1% done,” since most of the world doesn’t yet use AI or Claude Code.youtube+1
Lenny notes Anthropic’s large revenue and recent huge funding round; Boris credits user passion and feedback for continued improvement.
Lightning round highlights:
- Book recommendations: Functional Programming in Scala (technical elegance and types), Accelerando (captures accelerating change), The Wandering Earth short stories (Chinese SF perspective).youtube+1
- TV/movie: doesn’t watch much; liked the Three-Body Problem Netflix adaptation.
- Favorite products: Cowork (Chrome integration for automated web tasks) and the Acquired podcast.youtube+1
- Co‑work usage patterns: use it to clean desktops, summarize/respond to email, manage project status, connect multiple tools, and run many tasks in parallel while you step away.youtube+1
- Life motto: “Use common sense.” He sees many failures arising from blindly following process instead of first‑principles reasoning.
- Twitter/X presence: he started replying to Claude Code comments while on vacation, using it as a channel for bug reports and feature requests, which Claude helps him fix quickly.youtube+1
Closing: Boris invites people to tag him on X/Threads with bugs and feature requests; he emphasizes Claude Code’s growth as a function of user feedback.youtube+1

Pre‑2024: Anthropic develops a model roadmap: first excel at coding, then tool use, then computer use, as a safety‑driven progression.
Boris joins Anthropic (date not specified):
- Month 1: hacking prototypes to understand model capabilities.
- Month 2: post‑training work to understand “the layer under the layer.”
- Then returns to prototyping an agentic coding tool.
First prototype (“ClaudeCLI”): terminal-based agent that uses tools like bash. Surprise success: uses only a batch tool to tell Boris what music he’s listening to.
Internal demo posted; minimal initial reaction (two likes).
Late 2024: Claude Code grows as an internal tool at Anthropic; daily active users go “vertical” after tracking starts.
Several months of internal-only use: Claude Code used inside Anthropic for 4–5 months to study agent safety and behavior before external release.
February (around 2025): initial external release of Claude Code.
- Not an instant mainstream hit; early adopters “get it,” but broader users take months to adapt to the terminal-based agent.
2025:
- Mid‑2025: Anthropic holds its first developer conference (“code with Claude”); Boris predicts that by year’s end engineers may not need IDEs and AI will write all code, which shocks the audience.
- Opus 4 and Sonnet 4 launch; these “ASL3‑class” models dramatically increase Claude Code’s coding ability and usage, marking a major inflection point in growth.
- By May 2025: Claude Code writes perhaps 30% of Boris’s code; he still uses other tools (e.g., Cursor) heavily.
- November 2025: Claude Code crosses 100% of Boris’s coding; he stops editing lines by hand and ships 10–30 PRs daily via agents.youtube+1
2025–early 2026 – Rapid external adoption and ecosystem integration:
- SemiAnalysis report: Claude Code now authors ~4% of public GitHub commits, with a projection to reach ~20%; daily actives are still accelerating, with DAUs doubling in the previous month.youtube+1
- Spotify publicly states its best developers haven’t written a line of code since December due to AI help (including Claude).
- Claude Code expands from terminal to mobile apps, desktop app, web UI, IDE extensions, Slack, and GitHub integrations.
- Anthropic raises a large funding round at a valuation >$350B; Claude Code becomes a major revenue driver and multi‑billion‑dollar business line.youtube+1
Emergence of Cowork:
- Throughout 2025: Anthropic observes non‑technical and semi‑technical users misusing Claude Code for non‑coding tasks (SQL analysis, genome/MRI analysis, disk recovery, gardening, etc.), revealing latent demand for general agentic workflows.
- Data scientists at Anthropic start using Claude Code in the terminal for SQL, surprising Boris and indicating cross‑disciplinary adoption.
- A small team proposes embedding Claude Code into the desktop app with safer, more guided interactions; over ~10 days they build Cowork, using Claude Code to write the code, including a VM and security layers.youtube+1
- Cowork launches as a research preview and quickly becomes a hit, adopted more rapidly than early Claude Code.
February 19, 2026 – Interview date:
- Lenny notes the episode coincides roughly with Claude Code’s one‑year public anniversary.
- Boris states coding is now “largely solved” for his work, sees Claude increasingly proposing what to build, and expects this pattern to spread to other roles via Cowork and agents.youtube+1

3. Concise biography of Boris Cherny (as presented here)

Current role: Creator and head of Claude Code at Anthropic, leading the team building Anthropic’s agentic coding tools and closely involved with Cowork.youtube+1
Professional background:
- Before Anthropic, worked in big tech, including Meta/Facebook, where he was responsible for company‑wide code quality across large codebases (Facebook, Instagram, WhatsApp), focusing on improving developer productivity.
- At Meta, saw typical productivity gains of only a few percentage points per year despite large efforts, which informs his awe at the hundreds‑of‑percent gains from AI coding agents.
- Helped build Threads at Meta and enjoyed its product design, which is why he initially preferred it over Twitter/X.
- Briefly joined Cursor (a competing AI coding tool) due to admiration for the team and product, but returned to Anthropic after two weeks because he missed Anthropic’s safety‑driven mission.youtube+1
Technical profile:
- Self‑taught engineer; studied economics formally, began programming in middle school, initially to cheat on math tests via TI‑83 programs and solvers distributed to classmates.youtube+1
- Deep interest in functional programming and type systems; authored a book on TypeScript and founded a large TypeScript meetup, reflecting his appreciation for elegance in programming languages.
- Describes himself as historically a highly prolific coder, among the top few engineers at Instagram and now at Anthropic, even while leading a team.
- Currently codes almost entirely by orchestrating Claude Code agents rather than editing code directly, but still performs code review and supervision.youtube+1
Philosophy and product principles:
- Mission-driven: strongly motivated by aligning advanced AI with safety and social benefit; this mission was his primary reason for joining and re‑joining Anthropic.youtube+1
- Advocates for underfunding teams slightly to force heavy AI leverage, giving engineers unlimited tokens early, and only optimizing costs once an approach proves valuable.youtube+1
- Emphasizes “latent demand” as the central product lens, both for user behavior and for observing what the model itself tries to do when given tools.youtube+1
- Endorses Rich Sutton’s “bitter lesson”: bet on general, powerful models rather than over‑specialized systems; design for the capabilities of the next model rather than today’s limitations.youtube+1
Personal background and interests:
- Born in Odessa, Ukraine; emigrated as a child (family left in 1995).
- Grandfather was among the first programmers in the Soviet Union, using punch cards; this family history shapes his view of evolving programming abstractions.
- Lived for a period in rural Japan, where he was the only engineer and English speaker in town; formed neighborly ties by trading homemade miso and pickles.youtube+1
- Passionate about miso-making as a long‑timescale craft and sees it as his likely focus in a “post‑AGI” world.youtube+1
- Avid reader of sci‑fi and technical literature; frequently recommends Functional Programming in Scala, Accelerando, and Liu Cixin’s The Wandering Earth stories.youtube+1
- Enjoys the Acquired podcast and rarely watches TV; liked the Three-Body Problem series.youtube+1
- Personal motto: “Use common sense,” meaning to reason from first principles instead of blindly following process.

4. Memorable quotes by theme (with timestamps)

A. State of coding and future of roles

“Productivity per engineer has increased 200%.”
“At this point it’s safe to say that coding is largely solved. At least for the kind of programming that I do, it’s just a solved problem because Claude can do it.”
“I imagine a world, a few years in the future where everyone is able to program. And what does that unlock? Anyone can just build software anytime.”
“In some places the title ‘software engineer’ is going to start to go away and it’s just going to be replaced by ‘builder’ or maybe it’s just everyone’s going to be a product manager and everyone codes.”
“There’s always this question, should I learn to code? … My take is in a year or two it’s not going to matter.”youtube+1

B. Claude Code usage and agents

“At this point, 100% of your code is being written by AI. … 100% of my code is written by Claude Code. I have not edited a single line by hand since November.”youtube+1
“Every day I ship like 10, 20, 30 pull requests. Something like that every day.”
“I have never enjoyed coding as much as I do today because I don’t have to deal with all the minutia.”youtube+1
“The fun part is figuring out what to build and coming up with this. It’s talking to users. It’s thinking about these big systems. … And that’s what I get to do more of now.”youtube+1
“Coding now is describing what you want, not writing actual code.”

C. Product philosophy, latent demand, and building for the future

“Sometimes you find a thread, you just have to pull on it.”
“Latent demand is the idea that if you build a product in a way that can be hacked or misused to do something people want, that helps you learn where to take the product next.”
“When you see people abusing the product in this way … it’s just such a strong indicator that you should build a product for that and people are going to like it.”
“For Claude Code, we inverted it. We said the product is the model. We want to expose it. We want to put minimal scaffolding around it, give it the minimal set of tools, and let it decide.”
“From the very beginning, we bet on building for the model six months from now, not for the model of today.”

D. Safety, general models, and the “bitter lesson”

“There’s no road map for innovation. You just have to give people space and maybe the word is safety — psychological safety that it’s okay to fail.”youtube+1
“Don’t try to box the model in. … Almost always you get better results if you just give the model tools, give it a goal, and let it figure it out.”
“The more general model will always outperform the more specific model. … Over the long term, always bet on the more general model.”youtube+1
“Scaffolding can improve performance maybe 10–20%, but often these gains just get wiped out with the next model.”
“We call this the race to the top. … We released an open-source sandbox that works with any agent, not just Claude Code, because we wanted to make it really easy for others to do the same thing.”youtube+1

E. Organizational and management principles

“If you have a good idea you just really want to get it out there. No one has to force you to do that. That comes from you.”
“If you want to go faster, a really good way to do that is to just have Claude do more stuff.”youtube+1
“My advice generally is: don’t try to optimize, don’t try to cost cut at the beginning. Start by just giving engineers as many tokens as possible.”
“At small scale you’re not going to get a giant bill. The token cost is still relatively low compared to their salary.”
“On the Claude Code team, everyone codes. Our product manager codes, our engineering manager codes, our designer codes, our finance guy codes, our data scientist codes.”

F. Personal history, analogies, and worldview

“Programming has always changed in this way. … Before that it was punch cards. Before that it was switches. Before that it was hardware. Before that it was literally pen and paper.”
“The printing press is such a good analogy because this technology that was locked away to a small set of people … became accessible to everyone. It was inherently democratizing.”youtube+1
“I imagine a world a few years in the future where everyone is able to program. … But I do think in the meantime it’s going to be very disruptive and it’s going to be painful for a lot of people.”youtube+1
“For me, coding is very much a tool and it’s a way to do things. That said, not everyone feels this way.”
“Use common sense. … The best results that I see are people thinking from first principles and just developing their own common sense.”

Head of Claude Code: What happens after coding is solved | Boris Cherny

#podcast #ClaudeCode #Anthropic #AI-agents #software-engineering #Cowork
Source: Lenny’s Podcast – Feb 19, 2026 [page:1]

0. Metadata

Title:: Head of Claude Code: What happens after coding is solved | Boris Cherny
Show:: Lenny’s Podcast
Date:: 2026-02-19 [page:1]
Duration:: 1:27:45 [page:1]
Guest:: Boris Cherny (Head of Claude Code, Anthropic) [page:1]
Host:: Lenny Rachitsky [page:1]
Link:: https://www.youtube.com/watch?v=We7BZVKbCVw
Transcript:: https://www.lennysnewsletter.com/p/head-of-claude-code-what-happens [page:1]

1. High-level structure

2. Detailed outline (with timestamp blocks)

1. Cold open & setup (00:00–03:45)

“100% of my code is written by Claude Code”
- Claims he hasn’t edited a single line by hand since November; ships 10–30 PRs per day. [page:1, {ts:0}]
- Multiple agents running in parallel, even while recording. [page:1, {ts:5}]
Enjoyment and productivity
- Says he has never enjoyed coding as much because the minutiae are gone. [page:1, {ts:15}]
- States ~200% productivity-per-engineer increase at Anthropic since Claude Code. [page:1, {ts:1283}]
“Coding is largely solved” + everyone can program
- Asserts coding is largely solved, at least for his kind of programming. [page:1, {ts:24}]
- Envisions a world where everyone can program and can “just build software anytime.” [page:1, {ts:30}]
Lenny’s framing
- Episode coincides with roughly the one‑year anniversary of Claude Code. [page:1, {ts:61}]
- Notes Claude Code’s transformative impact on software engineering and Anthropic’s growth (round at >$350B valuation). [page:1, {ts:79}]

2. Cursor detour & return to Anthropic (03:45–08:41)

Two-week Cursor stint
- Joined Cursor ~6 months earlier, then returned to Anthropic after two weeks. [page:1, {ts:238}]
- Praises Cursor’s team and product; thought they saw AI coding’s future early. [page:1, {ts:259}]
Why he returned
- Realized he missed Anthropic’s safety mission—wanted to help shape the future of “this crazy thing.” [page:1, {ts:275}]
- Anthropic hallway test: ask anyone why they’re there; answer is always “safety.” [page:1, {ts:301}]
- Concludes mission-driven work is essential for his own happiness. [page:1, {ts:313}]

3. One year of Claude Code (05:35–13:29)

External impact metrics
- SemiAnalysis: ~4% of all GitHub commits authored by Claude Code; projected ~20% by end of year. [page:1, {ts:353}]
- Spotify headline: best developers haven’t written a line of code since December thanks to AI. [page:1, {ts:373}]
Growth pattern
- Finds 4% number “crazy,” notes real figure higher including private repos. [page:1, {ts:405}]
- More striking: acceleration—growth curves (DAU, etc.) steepening over time. [page:1, {ts:422}]
Anthropic labs strategy
- Model trajectory: great at coding → great at tool use → great at computer use, driven by safety considerations. [page:1, {ts:446}]
- Claude Code, MCP, and desktop app as “round one” outputs of the labs team. [page:1, {ts:469}]

4. Origin story & internal adoption (08:41–22:24)

Prototyping path
- Month 1 at Anthropic: hacking prototypes to map model capability boundaries. [page:1, {ts:526}]
- Month 2: post‑training work to understand “the layer under the layer” (model internals) to build better products. [page:1, {ts:543}]
- Returns to prototyping what becomes Claude Code. [page:1, {ts:571}]
“ClaudeCLI” prototype
- Early terminal tool “ClaudeCLI” demo: given only a bash tool, model figures out how to tell him what music he’s listening to. [page:1, {ts:582}]
- No explicit instructions; model autonomously decides how to use tools. [page:1, {ts:596}]
Lukewarm internal reception
- Internal demo post gets “two likes”; people expect IDE‑style coding tools, find terminal concept odd. [page:1, {ts:619}]
Under-resourcing as product lesson
- Built in the terminal because he was alone on the project; terminal was fastest way to keep up with rapidly improving models. [page:1, {ts:642}]
- Intentionally under-resourcing early helps force focus and adaptability. [page:1, {ts:653}]
Internal inflection
- After internal release, DAUs “go vertical”; Ben Mann nudges him to track DAUs early. [page:1, {ts:693}][page:1, {ts:700}]
External release
- External launch in February: not an immediate mainstream hit; early adopters “get it,” wider audience takes months. [page:1, {ts:710}][page:1, {ts:717}]
- Terminal interface is “surprising” and “alien,” demanding open‑minded users; now available in iOS/Android, desktop, web, IDEs, Slack, GitHub. [page:1, {ts:746}][page:1, {ts:758}]
Learning from users
- Describes the year as humbling; main signal is user feedback, and they “keep learning” from how people actually use it. [page:1, {ts:787}][page:1, {ts:799}]

5. Agents & 100% AI-written code; next frontier (16:17–32:15)

Innovation culture
- Innovation can’t be road-mapped; requires space, psychological safety to fail, and discipline to kill bad ideas quickly. [page:1, {ts:923}]
Claude Code’s share of Boris’s coding over time
- February: maybe ~20% of his code. [page:1, {ts:951}]
- May: ~30%; still using Cursor heavily. [page:1, {ts:951}]
- November: crosses 100%—no manual edits since then. [page:1, {ts:957}][page:1, {ts:1017}]
Current workflow
- 100% of his code written by Claude Code; ships 10–30 PRs per day. [page:1, {ts:979}][page:1, {ts:1007}]
- Still reviews code visually; Claude performs automatic code review on 100% of PRs with human reviews on top. [page:1, {ts:1024}][page:1, {ts:1033}]
Next frontier: ideas, not code
- Claude begins mining feedback, bug reports, telemetry to propose bug fixes and features—behaving more like a co‑worker. [page:1, {ts:1072}][page:1, {ts:1085}]
- Claims coding is “largely solved” for his programming; emerging frontier is adjacent tasks and general work via Cowork. [page:1, {ts:1097}][page:1, {ts:1112}]
Cowork usage examples
- Uses Cowork to pay parking tickets and handle team project management (spreadsheets, Slack, email). [page:1, {ts:1123}][page:1, {ts:1134}]
- Expects coding to become increasingly “solved” across stacks over coming months. [page:1, {ts:1147}]

6. Printing press analogy & “should I learn to code?” (27:55–36:01)

“Should I learn to code?”
- Today: users of Claude Code need some understanding of the underlying layer. [page:1, {ts:1953}]
- In 1–2 years: expects it won’t matter much whether you know how to code in the traditional sense. [page:1, {ts:1958}]
Printing press analogy
- Compares AI to Gutenberg’s printing press:
  - Before: <1% literacy; scribes handle reading/writing for illiterate elites. [page:1, {ts:1979}]
  - After press: printed material explodes; cost drops ~100x over 50 years. [page:1, {ts:2009}][page:1, {ts:2023}]
  - Over ~200 years, literacy climbs toward ~70%, enabling the Renaissance. [page:1, {ts:2039}]
Scribes and tedium
- Historical scribe reportedly pleased by the press because it removes tedious copying, freeing them for art and binding; Boris sees himself similarly freed from tedious coding. [page:1, {ts:2054}][page:1, {ts:2078}]
Long-run optimism vs short-run pain
- Envisions a world where everyone can program but stresses the transition will be disruptive and painful for many. [page:1, {ts:2417}][page:1, {ts:2429}]
- Argues this needs societal-level discussion; Anthropic has economists, policy, and social-impact staff engaged. [page:1, {ts:2283}]

7. Roles & careers in the AI era (36:01–46:34)

Next roles transformed
- Predicts AI will increasingly transform roles adjacent to engineering (PM, design, data science) and then “any kind of work you can do on a computer.” [page:1, {ts:2183}][page:1, {ts:2194}]
- Cowork as first agentic AI many non‑technical users experience. [page:1, {ts:2205}]
Definition of “agent”
- Clarifies an agent is an LLM that can use tools and act in your systems (Docs, email, commands), not just a chatbot. [page:1, {ts:2247}][page:1, {ts:2261}]
Disruption & jobs
- Acknowledges serious uncertainty about jobs; references Jevons paradox and stresses the need for industry and society to figure out how to handle it. [page:1, {ts:2273}][page:1, {ts:2301}]
Anthropic hiring & morale
- Claude Code team is hiring; personally enjoys coding more now and hears similar enthusiasm from many customers. [page:1, {ts:2326}][page:1, {ts:2339}]
Lenny’s informal polls
- ~70% of engineers and PMs say they enjoy work more with AI; ~10% less. [page:1, {ts:2682}]
- Designers: ~55% more, ~20% less—more ambivalence. [page:1, {ts:2701}]
Anthropic’s technical bar
- Even “non‑technical” roles go through technical interviews; many designers code and use the desktop app heavily. [page:1, {ts:2737}][page:1, {ts:2748}]
Advice: generalists + AI native
- Recommends:
  - Deeply experiment with AI tools; don’t be afraid of them. [page:1, {ts:2459}]
  - Become a generalist who crosses disciplines (e.g., product+infra, product+design, eng+business, eng+user-research). [page:1, {ts:2472}][page:1, {ts:2513}]
- On the Claude Code team, “everyone codes” (PM, EM, designer, finance, data scientist). [page:1, {ts:2494}]
Future of job titles
- Near term: PM/design/engineering still useful; ~50% overlap already. [page:1, {ts:2566}]
- By end of the year: expects titles to blur further; “software engineer” may vanish in favor of “builder,” or “everyone is a PM and everyone codes.” [page:1, {ts:2593}]

8. Latent demand & Cowork (46:34–54:04)

Principle of latent demand
- Defines latent demand as users misusing/hacking your product to do something it wasn’t designed for, revealing strong demand. [page:1, {ts:2852}][page:1, {ts:2859}]
Facebook examples
- Marketplace: discovered because ~40% of posts in Facebook groups were buy/sell. [page:1, {ts:2892}]
- Dating: ~60% of profile views were non-friends of opposite genders. [page:1, {ts:2931}]
Latent demand for Claude Code beyond coding
- Observes people using Claude Code for non‑technical tasks:
  - Gardening (growing tomatoes). [page:1, {ts:2962}]
  - Genome analysis, MRI analysis, photo recovery from corrupted drive (wedding photos). [page:1, {ts:2967}][page:1, {ts:2981}]
- Data scientists at Anthropic running SQL analysis from a terminal via Claude Code, despite its “low-level” feel. [page:1, {ts:2999}][page:1, {ts:3029}]
Latent demand from the model
- Extends the concept to “what the model is trying to do”:
  - Traditional approach: heavily orchestrated workflows where the model is only one component. [page:1, {ts:3072}][page:1, {ts:3082}]
  - Claude Code approach: treat the model as the product; minimal scaffolding; give tools and let it choose steps and order. [page:1, {ts:3092}][page:1, {ts:3097}]
Cowork build story
- Cowork emerges after the team sees non‑coding use of Claude Code and experiments for months. [page:1, {ts:3161}][page:1, {ts:3179}]
- Final idea: “take Claude Code and put it in the desktop app” with guardrails. [page:1, {ts:3190}]
- Built in ~10 days using Claude Code itself, including a full virtual machine and security scaffolding. [page:1, {ts:3197}][page:1, {ts:3214}]
- Released early as a research preview to learn both product-market fit and safety behavior. [page:1, {ts:3225}][page:1, {ts:3236}]

9. Safety & mechanistic interpretability (54:04–1:03:23)

Three layers of safety
1. Alignment & mechanistic interpretability:
  - Study neurons and internal activations; can detect concepts like deception and monitor them. [page:1, {ts:3280}][page:1, {ts:3297}]
2. Evals (“petri dish” lab tests):
  - Synthetic scenarios to test whether the model behaves safely. [page:1, {ts:3308}][page:1, {ts:3313}]
3. In-the-wild behavior:
  - Observe model behavior with real users in products like Claude Code and Cowork. [page:1, {ts:3319}]
Claude Code safety path
- Ran internally at Anthropic for 4–5 months before external release to study safety of one of the first widely used coding agents. [page:1, {ts:3331}][page:1, {ts:3337}]
Cowork safety path
- Similar process: strong alignment and eval performance, internal testing, limited customer pilots; then public research preview to observe real-world behavior. [page:1, {ts:3369}][page:1, {ts:3379}]
Mechanistic interpretability details
- References Chris Olah and mechanistic interpretability work; notes parallels between model neurons and biological neurons. [page:1, {ts:3440}][page:1, {ts:3458}]
- Mentions superposition, where multiple concepts are encoded across overlapping sets of neurons. [page:1, {ts:3498}][page:1, {ts:3502}]
Race to the top
- Anthropic open-sources safety tools like a sandbox for agents; works with any agent, not just Claude Code. [page:1, {ts:3550}][page:1, {ts:3557}]
- Goal: make it easy for other labs to adopt similar safeguards; “race to the top” on safety. [page:1, {ts:3543}]

10. Building AI products: tokens, scaffolding, bitter lesson (1:03:23–1:12:13)

Internal principles for the Claude Code team
- “What’s better than doing something? Having Claude do it.” [page:1, {ts:1451}]
- Underfund teams slightly so they’re forced to “claudify” workflows and automate more. [page:1, {ts:1466}][page:1, {ts:1472}]
- Encourage speed: if something can be done today, do it today; speed was their only advantage when the team was small. [page:1, {ts:1498}][page:1, {ts:1517}]
Unlimited tokens as a perk
- Advises CTOs not to optimize token costs early; give engineers as many tokens as possible to prototype ambitious ideas. [page:1, {ts:1563}][page:1, {ts:1593}]
- Argues small-scale experiments won’t yield huge bills compared to salary; optimize once something works and usage scales. [page:1, {ts:1629}][page:1, {ts:1648}]
Don’t box the model in
- Critiques rigid workflows where the model must follow hard-coded steps. [page:1, {ts:3832}][page:1, {ts:3838}]
- Recommends: give the model tools and a goal, let it decide steps and retrieve context; scaffolding that over-orchestrates is less necessary now. [page:1, {ts:3826}][page:1, {ts:3868}]
Rich Sutton’s “bitter lesson”
- Summarizes: more general models outperform specific models over time. [page:1, {ts:3885}]
- Corollaries:
  - Don’t overinvest in tiny models and heavy fine-tuning where general models can be used. [page:1, {ts:3909}][page:1, {ts:3921}]
  - Scaffolding may yield 10–20% performance boosts, but those are often wiped out by the next model generation. [page:1, {ts:3936}][page:1, {ts:3942}]
Build for the model six months out
- Claude Code was designed for a future where models could write most of the code, even when early models couldn’t. [page:1, {ts:3956}][page:1, {ts:3981}]
- Inflection occurred with Opus 4 and Sonnet 4; Opus 4 as first ASL3‑class coding model; usage and growth went exponential. [page:1, {ts:4001}][page:1, {ts:4009}]
- Advice to startups: accept weak product-market fit for 6 months while building for future model capabilities; then “hit the ground running” when new models drop. [page:1, {ts:4020}][page:1, {ts:4032}]
Predictable improvement dimensions
- Expects:
  - Better tool/computer use over time. [page:1, {ts:4064}]
  - Much longer unattended runs (from ~15–30 seconds with earlier Sonnet to 10–30 minutes or longer with Opus 4.6). [page:1, {ts:4075}][page:1, {ts:4099}]

11. Pro tips & usage patterns (1:08:38–1:21:00)

No single “right” workflow
- Emphasizes that developers differ; suggests asking Claude Code itself how to configure and use it. [page:1, {ts:4135}][page:1, {ts:4151}]
Tip 1: Use the best model + max effort
- Recommends Opus 4.6 with maximum effort enabled. [page:1, {ts:4157}]
- Argues cheaper models often consume more tokens overall due to needing more corrections and handholding. [page:1, {ts:4168}][page:1, {ts:4178}]
Tip 2: Use plan mode
- Starts ~80% of tasks in plan mode; injected instruction “please don’t write any code yet.” [page:1, {ts:4197}][page:1, {ts:4204}]
- Terminal shortcut: Shift+Tab twice; plan mode also available in desktop, web, and soon mobile; launched in Slack integration. [page:1, {ts:4214}][page:1, {ts:4225}]
- Once plan looks good, he auto-accepts edits; with Opus 4.6, often one-shots tasks correctly. [page:1, {ts:4231}][page:1, {ts:4236}]
Tip 3: Explore different interfaces
- Suggests trying terminal, desktop app, mobile apps, Slack, etc.; stresses it’s the same agent behind each. [page:1, {ts:4243}][page:1, {ts:4263}]
Cowork usage patterns
- Start by:
  - Having it use tools (e.g., clean desktop, summarize/respond to email). [page:1, {ts:4930}][page:1, {ts:4945}]
  - Connecting tools (email, Slack, spreadsheets) so it can coordinate workflows like weekly status pings. [page:1, {ts:4951}][page:1, {ts:4966}]
  - Running many tasks in parallel while you step away; he kicks off tasks then gets coffee. [page:1, {ts:4976}][page:1, {ts:4985}]

12. Biography & philosophy threads (throughout)

Career background
- Formerly led code quality at Meta across large codebases (Facebook, Instagram, WhatsApp), focused on developer productivity. [page:1, {ts:1294}][page:1, {ts:1305}]
- Used to see only low single-digit annual productivity gains despite large efforts; contrasts that with current hundreds-of-percent gains. [page:1, {ts:1311}][page:1, {ts:1322}]
- Helped build Threads; appreciates its clean design; used it exclusively before becoming more active on Twitter. [page:1, {ts:5108}]
Self-taught engineering
- Studied economics, taught himself programming in middle school. [page:1, {ts:1693}][page:1, {ts:1699}]
- First serious program: math test cheating on TI‑83; later built equation solvers and shared via cables with classmates; the whole class got As and was caught. [page:1, {ts:1716}][page:1, {ts:1745}]
Relationship to code
- Sees programming as a practical tool to build things, not an end in itself, though he appreciates the beauty of functional programming and type systems. [page:1, {ts:1751}][page:1, {ts:1807}]
- Wrote a book on TypeScript and founded a large TypeScript meetup; heavily influenced by functional programming in Scala. [page:1, {ts:1765}][page:1, {ts:4591}]
Historical view on programming abstractions
- Places AI-assisted coding on a continuum from pen-and-paper computers to switches, punch cards, and virtual machines. [page:1, {ts:1852}][page:1, {ts:1878}]
Personal background
- Born in Odessa, Ukraine; family emigrated in 1995; expresses deep gratitude for life in the US. [page:1, {ts:3756}][page:1, {ts:3779}]
- Grandfather among the first Soviet programmers, using punch cards; childhood stories of stacks of cards at home. [page:1, {ts:3706}][page:1, {ts:3719}]
Time in rural Japan
- Lived in a small town as the only engineer and English speaker; life organized around seasons and local agriculture. [page:1, {ts:4355}][page:1, {ts:4735}]
- Developed a love for miso-making and pickling; trades miso with neighbors; miso teaches him to think on multi-year timescales. [page:1, {ts:4385}][page:1, {ts:4408}]
- States that post-AGI he’d likely make miso; sees it as a long-horizon craft counterbalancing high-speed AI work. [page:1, {ts:4419}][page:1, {ts:4442}]
Life motto
- “Use common sense”: warns against blindly following process or momentum without first-principles reasoning; encourages trusting when something “smells weird.” [page:1, {ts:5054}][page:1, {ts:5073}]

13. Lightning round (1:14:02–end)

Book recommendations
- Functional Programming in Scala – favorite technical book; elegant types and FP mindset. [page:1, {ts:4591}][page:1, {ts:4608}]
- Accelerando (Charles Stross) – fast-paced SF capturing the feel of accelerating change and approaching singularity. [page:1, {ts:4625}][page:1, {ts:4650}]
- The Wandering Earth (Liu Cixin) – short-story collection; loves Chinese SF perspective. [page:1, {ts:4662}][page:1, {ts:4679}]
TV/movie
- Rarely watches TV; praises Netflix’s Three-Body Problem adaptation. [page:1, {ts:4799}][page:1, {ts:4811}]
Favorite products
- Cowork (especially Chrome integration) – pays fines, cancels subscriptions, fills out forms, handles tedious tasks. [page:1, {ts:4830}][page:1, {ts:4885}]
- Podcast: Acquired (Ben Gilbert & David Rosenthal), recommends Nintendo episode. [page:1, {ts:4857}][page:1, {ts:4868}]
Cowork eval anecdote
- Lenny’s earlier post of “50 non‑technical uses of Claude” is used as an eval; Cowork could accomplish ~48/50 before launch. [page:1, {ts:5011}][page:1, {ts:5022}]
Social media usage
- Was primarily on Threads; started using Twitter while bored on a coding vacation in Europe, responding to Claude Code tweets and bug reports. [page:1, {ts:5108}][page:1, {ts:5149}]
- Uses Twitter as a feedback channel; surprised users with speed of bug fixes via Claude Code. [page:1, {ts:5161}][page:1, {ts:5177}]
Call to action
- Invites listeners to tag him on Threads/Twitter with bugs and feature requests; credits user feedback as main driver of Claude Code’s improvement. [page:1, {ts:5223}][page:1, {ts:4554}]

3. Claude Code / Cowork timeline (for separate note linking)

Pre‑2024 – Strategy: coding → tools → computer use, driven by safety. [page:1, {ts:446}]
Early Anthropic tenure – Month 1 prototypes, Month 2 post‑training / model-understanding. [page:1, {ts:526}][page:1, {ts:543}]
“ClaudeCLI” prototype – terminal-only agent using a bash tool to infer what music he’s listening to. [page:1, {ts:582}][page:1, {ts:596}]
Internal launch – lukewarm reaction initially (two likes); later internal DAU inflection. [page:1, {ts:619}][page:1, {ts:700}]
4–5 months internal-only – used to study agent safety before external release. [page:1, {ts:3331}]
February external launch – early adopters only; takes months for broad understanding. [page:1, {ts:710}][page:1, {ts:717}]
May 2025 – developer conference (“code with Claude”); predicts end-of-year no-IDE coding; room gasps. [page:1, {ts:839}][page:1, {ts:852}]
Opus 4 / Sonnet 4 release – ASL3-class coding; usage and growth inflect hard. [page:1, {ts:4001}][page:1, {ts:4009}]
November 2025 – 100% of Boris’s code written by Claude Code; no manual edits. [page:1, {ts:979}][page:1, {ts:1017}]
By early 2026 – ~4% of public GitHub commits by Claude Code, DAUs doubling month-over-month. [page:1, {ts:353}][page:1, {ts:405}]
Cowork – built in ~10 days, fully implemented by Claude Code; launched as research preview, quickly becomes a hit. [page:1, {ts:3197}][page:1, {ts:3161}]

(You can transclude this block into a Claude Code Timeline note.)

4. Quote blocks (Obsidian-ready, with timestamps)

4.1 Coding and roles

100% of my code is written by Claude Code. I have not edited a single line by hand since November. [page:1, {ts:0}]

Every day I ship 10, 20, 30 pull requests. [page:1, {ts:5}]

I have never enjoyed coding as much as I do today because I don’t have to deal with all the minutia. [page:1, {ts:15}]

At this point it’s safe to say that coding is largely solved. At least for the kind of programming that I do, it’s just a solved problem because Claude can do it. [page:1, {ts:1097}]

I imagine a world a few years in the future where everyone is able to program. Anyone can just build software anytime. [page:1, {ts:2417}]

In some places the title “software engineer” is going to start to go away and it’s just going to be replaced by “builder,” or maybe it’s just everyone’s going to be a product manager and everyone codes. [page:1, {ts:2593}]

4.2 Product principles and latent demand

Sometimes you find a thread, you just have to pull on it. [page:1, {ts:974}]

Latent demand is this idea that if you build a product in a way that can be hacked or misused to do something people want to do, then this helps you as the product builder learn where to take the product next. [page:1, {ts:2859}]

When you see people abusing the product in this way, using it in a way that it wasn’t designed in order to do something that is useful for them, it’s just such a strong indicator that you should build a product for that and people are going to like it. [page:1, {ts:3039}]

For Claude Code, we inverted it. We said the product is the model. We want to expose it. We want to put minimal scaffolding around it, give it the minimal set of tools, and let it decide. [page:1, {ts:3092}]

From the very beginning, we bet on building for the model six months from now, not for the model of today. [page:1, {ts:3956}]

4.3 Safety, bitter lesson, and building with agents

Don’t try to box the model in. Almost always you get better results if you just give the model tools, give it a goal, and let it figure it out. [page:1, {ts:3826}]

The more general model will always outperform the more specific model. Over the long term, always bet on the more general model. [page:1, {ts:3885}][page:1, {ts:3916}]

Scaffolding can improve performance maybe 10–20%, but often these gains just get wiped out with the next model. [page:1, {ts:3936}]

Start by just giving engineers as many tokens as possible. [page:1, {ts:1563}]

If you want to go faster, a really good way to do that is to just have Claude do more stuff. [page:1, {ts:1521}]

4.4 Historical analogies and worldview

Programming has always changed in this way. Before that it was punch cards. Before that it was switches. Before that it was hardware. And before that it was literally pen and paper. [page:1, {ts:1852}][page:1, {ts:1872}]

The printing press is such a good analogy because this technology that was locked away to a small set of people became accessible to everyone. It was inherently democratizing. [page:1, {ts:1979}][page:1, {ts:2366}]

We couldn’t be talking today if the printing press hadn’t been invented. Our microphones wouldn’t exist. None of the things around us would exist. [page:1, {ts:2401}]

I imagine a world a few years in the future where everyone is able to program. And what does that unlock? Anyone can just build software anytime. And I have no idea. [page:1, {ts:2417}]

4.5 Personal philosophy / meta

For me, coding is very much a tool and it’s a way to do things. [page:1, {ts:1807}]

Use common sense. I think a lot of the failures that I see, especially in a work environment, is people just failing to use common sense. [page:1, {ts:5054}]

The best results that I see are people thinking from first principles and just developing their own common sense. [page:1, {ts:5073}]

5. Suggested Obsidian links and tags

5.1 Candidate notes to create/link

People/Boris Cherny → link to Boris Cherny biography section. [page:1]
Concepts/Latent demand (product) → link from product sections. [page:1]
Concepts/Agents and tool use → link from agent discussion. [page:1]
Concepts/Bitter lesson (Sutton) → link from model-generalization section. [page:1]
Products/Claude Code and Products/Cowork → transclude timeline and tips. [page:1]

You can now split this into multiple notes (e.g., /People/Boris Cherny, /Products/Claude Code, /Concepts/Latent demand) and use block refs to the quote blocks and outline sections as needed.