Jonathon's AI Wiki

Boris Cherny on Lenny's Podcast — What Happens After Coding Is Solved

11 min read

Source: raw/Head_of_Claude_Code_-_What_happens_after_coding_is_solved_Boris_Cherny.md (Lenny’s Podcast, “Head of Claude Code: What happens after coding is solved,” Boris Cherny, 1:27:45, uploaded 2026-02-19, YouTube We7BZVKbCVw). Product names normalized from auto-captions per wiki convention (“Quad/Cloud Code” → Claude Code, “Cower/Coda/co-work” → Cowork, “Entropic” → Anthropic). Model-version references (“Opus 4.6 is the best model”) are accurate as of the Feb 2026 recording — Opus 4.7 shipped later.

The full interview behind the @cyrilXBT promo post that triggered the Boris Cherny entity page. Boris is the creator and Head of Claude Code at Anthropic. Stripped of the influencer framing, the actual conversation is a dense hour-plus on how he personally works (100% Claude-written code, 10-30 PRs/day, ~5 agents running), the principles he codifies for his team, how to build AI products, three concrete Claude Code pro tips, and his “coding is virtually solved / printing-press” thesis for what comes next.

Key Takeaways

  • “100% of my code is written by Claude Code. I have not edited a single line by hand since November.” Ships 10-30 PRs every day; usually has ~5 agents running at once (literally during the recording). Splits his time roughly a third terminal / a third desktop app / a third iOS app — “I did not think this would be the way I code in 2026.”
  • He still reads the code, and Claude reviews 100% of PRs at Anthropic with a human-review layer after — except pure throwaway prototype code. Verification didn’t disappear; it moved.
  • Productivity per engineer at Anthropic is up 200% (in PRs) over the year while the team ~4x’d. For contrast, at Meta a year of hundreds of engineers on code-quality moved productivity “a few percentage points.”
  • The most actionable section is three Claude Code pro tips (best model + max effort, plan mode for ~80% of tasks, try every interface) — see below. Boris’s caveat: “there’s no one right way to use Claude Code” and you can ask Claude Code about itself (it can edit your settings).
  • His mental model for the whole shift is the printing press, and his one-line life motto is “use common sense.”

His Daily Workflow

  • 100% Claude-written since November, 10-30 PRs/day. Self-describes as a prolific coder even as team head (was a top-few engineer at Instagram).
  • ~5 agents running continuously. Starts a batch on waking; checks them from the Claude iOS app, code tab. Says he hasn’t really felt “agent anxiety” precisely because he always has agents running and isn’t locked into a terminal.
  • Thirds: ~⅓ terminal, ⅓ desktop app, ⅓ iOS app. “Coding now is describing what you want, not writing actual code.”
  • Worked example — the memory leak: a newer engineer simply told Claude Code “there’s a leak, can you figure it out?” Claude took a heap snapshot, wrote its own just-in-time analysis tool, found the issue, and opened a PR — faster than Boris doing it the traditional (manual debugger) way. His lesson: veterans get stuck in old-model habits — “it’s not Sonnet 3.5 anymore.”

Three Claude Code Pro Tips

Boris’s explicit caveat first: “There’s no one right way to use Claude Code” — it’s a dev tool, devs differ, find your own path. And you can just ask Claude Code — “it kind of knows about itself, it can edit your settings, it can make recommendations.”

  1. Use the most capable model, with maximum effort always on. Currently (Feb 2026) Opus 4.6 + max effort. Counterintuitive cost point: a cheaper model (Sonnet) can take more tokens to finish the same task because it needs more correction/hand-holding — “it’s actually not obvious that it’s cheaper if you use a less expensive model.” The capable model is often cheaper and less token-intensive per completed task.
  2. Use plan mode — he starts ~80% of tasks in plan mode. Demystified: “all it is is we inject one sentence into the model’s prompt: ‘Please don’t write any code yet.’ That’s it.” Shift+Tab twice (terminal), a button (desktop/web), launched for Slack, coming to mobile. After a good plan he auto-accepts edits — “if the plan works good, it’s just going to one-shot it” with Opus 4.6.
  3. Play around with different interfaces. Terminal, iOS/Android, desktop, Slack, GitHub — “it’s the same Claude agent running everywhere.” Don’t assume Claude Code = terminal. Running many in parallel = “multi-clauding” (designers use the desktop app’s code tab for this).

Principles for Building AI Products

  • Latent demand — “the single most important principle in product.” When people “abuse” a product to do something it wasn’t designed for, build the special-purpose product. Examples he cites: Facebook Marketplace (40% of group posts were buy/sell), Facebook Dating (60% of profile views were opposite-gender non-friends), and Cowork (people used Claude Code to grow tomatoes, analyze their genome, recover wedding photos from a corrupt drive, read an MRI; data scientist Brendan ran SQL in the terminal → the next week every data scientist did).
  • Modern latent demand = watch what the model is trying to do, and make that easier (research term: “being on distribution”). Claude Code “inverted” the usual pattern — “the product is the model”: minimal scaffolding, minimal tools, let it choose which tools to run and in what order.
  • Don’t box the model in. Avoid rigid step-1→step-2→step-3 orchestrators. “Almost always you get better results if you just give the model tools, give it a goal, and let it figure it out.” A year ago you needed scaffolding; now you mostly don’t.
  • Don’t front-load context — give it a tool to fetch context. “Don’t try to give it a bunch of context up front. Give it a tool so that it can get the context it needs.”
  • The bitter lesson (Rich Sutton). The more general model always beats the more specific one — bet on the general model, avoid tiny models / fine-tuning where you can. Scaffolding buys ~10-20% “but often these gains just get wiped out with the next model — it’s almost better to just wait for the next one.”
  • Build for the model 6 months out, not today’s model. Claude Code’s founding bet. Uncomfortable (PMF is poor for ~6 months) but when the model lands you “hit the ground running.” Predictable improvement axes: better tool/computer use, and longer unattended runtime (Sonnet 3.5 ran ~15-30s before going off the rails; Opus 4.6 runs 10-30 min unattended, sometimes hours/days/weeks).
  • Release early — to discover latent demand, and (at Anthropic) to study safety in the wild. Claude Code ran internally 4-5 months before launch.

How He Runs the Team

  • “What’s better than doing something? Having Claude do it.” Codified for new joiners — “Claudify” the work.
  • Under-resource on purpose. Putting one engineer on a project forces Claudification. “You will get more out of the AI tooling if you have fewer people working on something” — empower great engineers and they figure it out.
  • Go faster — if you can do it today, do it today. Early on, speed was Claude Code’s only advantage in a crowded market; still a core team value.
  • Give engineers as many tokens as possible; don’t cost-optimize early. “Start by giving engineers as many tokens as possible.” Optimize (Haiku/Sonnet vs Opus) only after an idea works and scales. Some Anthropic engineers now spend hundreds of thousands/month in tokens. An individual experimenting is “still probably relatively low relative to their salary.”

Cowork — Agentic AI Beyond Coding

Same Claude agent, in the desktop app’s Cowork tab, pointed at non-coding tasks. Boris uses it daily — paid a parking ticket, runs all of his team’s project management (a single team status spreadsheet; every Monday Cowork messages every engineer on Slack who hasn’t filled in their status). Born from latent demand and growing faster than Claude Code did early. (Lenny’s “50 non-technical use cases” post was used by a PM as an eval gate — Cowork clearing 48/50 was the “okay, it’s good” bar.)

His recommended Cowork onboarding:

  1. Have it use a tool — clean up your desktop, summarize your email, respond to your top three emails.
  2. Connect tools — “look at my top emails and send Slack messages” / put them in a spreadsheet.
  3. Run a bunch in parallel — kick off several tasks, “then I just go get a coffee while it runs.”

The Big Picture

  • “Coding is virtually solved” — at least for the kinds of programming he does — and over the next few months becomes “increasingly solved” across codebases/stacks. Code review became the next bottleneck; now humans are needed for what to build / prioritize — and Claude is starting to help there too (reading feedback, bug reports, telemetry → proposing fixes and PRs, “a little more like a co-worker”).
  • Next roles impacted: adjacent-to-engineering first — PM, design, data science — then any computer-based work. “By the end of the year the title software engineer is going to start to go away… replaced by builder, or everyone’s a product manager and everyone codes.” He already sees ~50% role overlap.
  • Printing-press analogy (his core frame): mid-1400s literacy was <1% (scribes); the 50 years after Gutenberg produced more printed material than the prior 1,000, cost down ~100x; literacy reached ~70% globally over 200 years. Programming democratizes the same way — “anyone can build software anytime” — disruptive and painful in the interim; an explicitly societal conversation. The scribe who welcomed the press because he preferred art + bookbinding to copying = how Boris feels about shedding tedious coding.
  • Should you learn to code? If you use agents today, you still need to understand “the layer under” — “but in a year or two it’s not going to matter.”
  • Be a generalist. On the Claude Code team everyone codes (PM, EM, designer, finance, data scientist); the strongest engineers cross disciplines (product+infra, product+design, business sense, user empathy). Generalists who cross domains will be rewarded most.

Safety (Anthropic Framing)

  • The coding → tool use → computer use trajectory is how Anthropic thinks about building safe AGI — each stage is more capable and more to study.
  • Three layers of safety study: (1) alignment + mechanistic interpretability (monitoring neurons — e.g., a “deception” concept; superposition), (2) evals (petri-dish synthetic situations), (3) behavior in the wild (why they release early). “Race to the top”: Anthropic open-sourced a Claude Code sandbox that works with any agent.

Notable Quotes

  • “100% of my code is written by Claude Code. I have not edited a single line by hand since November.”
  • “There’s no one right way to use Claude Code… luckily, you can ask Claude Code — it kind of knows about itself.”
  • On plan mode: “All it is is we inject one sentence into the model’s prompt — ‘Please don’t write any code yet.’ That’s it.”
  • “Don’t try to give it a bunch of context up front. Give it a tool so that it can get the context it needs.”
  • “Build for the model 6 months from now, not for the model of today.”
  • Life motto: “Use common sense… if something smells weird, it’s probably not a good idea.”
  • Post-AGI plan: “I’d probably be making miso.”

Try It

  • Adopt the three tips today: set the most capable model + max effort, start ~80% of tasks in plan mode (Shift+Tab twice), and try Claude Code outside the terminal (desktop/iOS/Slack). Cross-check against Opus 4.7 Best Practices and Picking the Right Model (his “capable model is often cheaper per outcome” claim is the same thesis).
  • Run the “ask Claude Code about itself” move: have it review your CLAUDE.md and settings and recommend changes — pairs with Anthropic’s Best Practices.
  • Try the latent-demand lens on your own product/work: where are people “abusing” an existing tool to do something it wasn’t built for? Build the special-purpose version.
  • Set up the Cowork project-management pattern: one team status spreadsheet + a Monday routine that pings whoever hasn’t filled it in. See Claude Cowork.

Graph View

Backlinks