Claude Code
Anthropic's agentic coding tool that lets Claude work on code end to end.
Claude Code grew from a terminal hack into about 4% of public GitHub commits, with growth jumping when Opus 4 and Opus 4.5 landed. It writes 100% of Boris Cherny's code, and runs went from 15-30 seconds to 20-30 minutes. Lesson: build for the model six months out, give it tools and a goal over rigid workflows, ship rough research previews (Cowork took about 10 days), and watch how users misuse the product.
- External releaseFelt alien; ~20% of Cherny's code
- Opus 4 shipsGrowth takes off; 30% of code
- Opus 4.5 shipsEnd-to-end agent work; 100% of code
- Opus 4.64% of public GitHub commits
From Boris Cherny and Dianne Penn on podcasts
What was said about Claude Code
1From terminal hack to 4% of GitHub commits
Claude Code began as a test of what the model could do with one tool. Boris Cherny spent a month on prototypes and a month on post-training. With only a bash tool, the model answered "what music am I listening to" without being told how.
It was incubated in Anthropic Labs alongside MCP. His internal announcement got two likes, and the terminal stayed because it was the easiest thing for one person to build and no other form could keep pace with the model. Internal daily users went vertical, but the external release in February took months to catch on because it felt alien.
The model and the product lifted each other's adoption. The team felt Claude Code's magic for months before Opus 4.5. Growth took off with Opus 4 in May and again in November, when Opus 4.5 let users run work end to end in an agent manner.
Dianne Penn: "you need frontier products in order to have frontier models and for people to feel the magic of frontier models." Opus 4.5 landed during winter break and drove the Claude Code inflection.
Labs runs big bets with tiny pods and loosely held prototypes. Labs keeps pods small, sometimes one engineer, pushes each idea toward its 10x, 100x or 1000x version, and turns off bets to revisit them a model generation or two later. Claude Code, Skills, MCP and Claude Design came out of it.
Penn: "you can be very strongly held opinion about the theme or the area and then more weekly held about the exact prototype." Claude Code was under-resourced on purpose, and Cherny says underfunding pushes engineers to automate the rest with Claude.
Usage has reached a large share of the world's code. Claude Code reached 4% of public GitHub commits, daily active users doubled in a month, and it writes 100% of Cherny's code: about 20% in February, 30% in May, 100% by November. He ships 10 to 30 pull requests a day.
Runs grew from 15 to 30 seconds to 20 to 30 minutes unattended with Opus 4.6. Anthropic engineers ship about 8x more code per quarter than in 2025, though that counts volume, not quality. Agentic coding did not exist a year and a half ago and is already bigger than earlier AI coding use.
2Building for the next model, not today's
Design for where models will be six months out. Cherny bet early that models would write most code even when Claude Code wrote little of his. When Opus 4 and Sonnet 4 shipped, growth went exponential.
"It's going to be uncomfortable cuz your product market fit won't be very good for the first 6 months." Cat Wu adds the counterweight: building for a super-strong model is easy, and the hard part is eliciting maximum capability from today's model.
Prototype features that don't work yet, then drop in each new model. Code review was attempted several times and only became reliable enough to gate merges with Opus 4.5, 4.6 and Sonnet 4.6, running multiple review agents in parallel.
Anthropic's engineering team now relies on Claude code review passing before merging PRs. Building early shows exactly what is missing for the product to work.
Give the model tools and a goal, not a fixed workflow. Claude Code was built with minimal scaffolding: the model picks which tools to run, in what order, and fetches its own context. Scaffolding may add 10 to 20%, but the next model often wipes that out.
Cherny frames this as the modern form of latent demand: look at what the model is trying to do and make that easier, which researchers call staying on distribution.
Delete harness crutches each time a new model ships. The to-do list was added because early Claude stopped after changing 5 of 20 call sites in a refactor. With Opus 4 and later, the model used it unprompted.
At each launch the team reads the whole system prompt and removes sections the model no longer needs. New models also unlock features that were not accurate enough before.
Retest what you assume the model can't do. A newer engineer asked Claude Code to find a memory leak; it took the heap snapshot, wrote itself an analysis tool and opened a PR faster than Cherny debugging by hand.
Fiona Fung saw engineers resist AI tools after early mistakes in Sonnet 3.5 or 3.6 without accounting for the rate of improvement. Cherny: "it's not Sonnet 3.5 anymore."
Run the most capable model; it is often cheaper overall. Cherny runs Opus 4.6 at maximum effort always, because a less capable model like Sonnet often burns more tokens on correction and hand-holding.
He starts about 80% of tasks in plan mode, which injects one sentence telling the model not to write code yet, then auto-accepts edits; with a good plan Opus 4.6 usually one-shots it.
3How the team ships in days
- Feature timeline: 6 months
- Timeline: 1 month
- Sometimes 1 week or 1 day
- Announcement out next day
Features ship in a week or two
From Claude Code team on podcasts
Label features as research previews to lower the cost of shipping. Almost every feature ships branded as a research preview, so it can go out in a week or two. Feature timelines fell from 6 months to 1 month, sometimes 1 week or 1 day.
Claude Code itself was used internally for 4 or 5 months before release, and Cowork shipped rough. The shipping pace is cited outside Anthropic as the textbook case of survival through speed.
An evergreen launch room turns announcements around the next day. Once an engineer has dogfooded a feature, they post it in a standing launch room, and docs, product marketing and DevRel turn around the announcement the next day.
The PM's job is to build this process and define when partners get pulled in. Cat Wu says Mythos helps only a little; process and expectations explain the pace.
Write a goal sharp enough to rule options out. LLMs are so general that ambiguity is the default. One goal: professional developers at enterprises safely reach zero permission prompts, targeting permission fatigue.
PRDs shrink to a one-pager on goals, delightful use cases and current failure modes, kept for ambiguous or infrastructure-heavy work.
Prefer Claude doing the work to doing it yourself. The team's rule is "what's better than me doing it? Claude doing it." Some projects get a single engineer who ships fast and automates the rest.
There is no roadmap; teams are free to fail even if 80% of ideas are bad, and bad ideas get cut quickly.
Pair high agency with high accountability and outcome metrics. Anyone can propose and build, then must state the hypothesis and the accountability. Each team defines "bad" (unrecoverable, like a crash) and "sad" (recoverable pain, like flicker) thresholds.
Fung treats token usage and lines of code as motion, not progress: "Don't forsake motion for progress."
4Code review and verification
Commit your definition of good so Claude Code Review can check it. Human review was a major bottleneck until Claude Code Review, which did not exist the year before. Specs and quality scales checked into the repo let it confirm changes still match.
Claude reviews 100% of PRs at Anthropic, followed by a human layer; only throwaway prototype code skips human review. Experts still review areas needing deep subject knowledge, on a trust but verify basis.
The source code leak was a process failure, not a model failure. A person working with Claude on a PR changed how packages are released, and the change passed two layers of human review.
The person is still at Anthropic, and most added safeguards have already shipped.
Going on vibes instead of evals is contested. Cherny said the team runs no evals and goes on vibes, which sparked a debate; Kiriti Badam rejects the idea of an eval set you could bet your life on.
5Cowork: built from latent demand in Claude Code
Cowork came from non-coders bending Claude Code to other tasks. People used Claude Code for tomato plants, genome analysis and recovering wedding photos, and data scientists ran SQL in the terminal. Cowork was built in about 10 days by Claude Code inside the desktop app, including a virtual machine and guardrails.
It shipped rough as a research preview. Claude for small business followed the same way, bundling plugins into a toggle after owners kept asking. Cherny calls latent demand "the single most important principle in product."
Adopt Cowork in stages: one tool, then connected tools, then parallel runs. Start with one tool, like cleaning your desktop, then connect tools, then run many tasks in parallel. Cherny has it message every engineer who hasn't filled the Monday status sheet.
A PM judged Cowork ready once it handled 48 of Lenny's 50 non-technical Claude Code use cases.
Connect Slack, Calendar, Gmail and Drive before asking for real work. With context connected and a narrative, drafts and template provided, Cowork worked for about an hour and produced a 20-page conference deck overnight, needing one feedback round.
Ask for an outline first and make the decisions yourself. A restaurant owner pointed it at a cluttered folder; it found her menus and produced a local price comparison.
6How the team's leaders run their work through Claude
A Claude session on every repo gives a manager full visibility. Fung runs a remote Claude Code session across all repos with access to Slack and team metrics, and reviews each month what shipped, how it did and which incidents followed.
It pulls themes across incidents and flags quality hot spots. "I actually have these sessions to enable me to have conversations with folks that I support."
Replace manual feedback triage with an overnight routine that drafts PRs. Fung's routine summarizes feedback channels, kicks off agents on polish fixes and leaves PRs ready by morning.
Early on, Cherny fixed each internal feedback item within minutes so users felt heard; now he points Claude at the channel and reviews its PRs.
Scheduled Cowork tasks cover dashboards, misalignment and coaching. Amol Avasare has Cowork review 20 to 25 charts each morning via the Chrome extension and MCP, and scan Slack weekly for misalignment. A colleague's scan surfaced major misalignment that would have caused duplicated work.
He tracks false positives and negatives to build trust in what it flags. He also asks Claude to play his manager weekly using her writing and his Slack activity; he calls the hit-or-miss result "soft coaching."
7Who the team hires and how roles merge
Hire engineers with product taste instead of more PMs. The team favors engineers who go from user feedback to a shipped feature in a week with almost no PM involvement; nearly all PMs and designers have shipped code.
On the team the PM, designer, finance lead and data scientist all code, and Cherny sees about 50% role overlap. Claude Code's PMs are engineers, making it the exception to Anthropic's squeezed-PM problem. Cherny sets the 3 to 6 month vision and Wu plans the route.
Pair product builders with deep systems experts. Early Claude Code had strong product generalists but lacked distributed systems expertise.
Mobile expertise is still needed, but in smaller numbers because people flex across platforms.
Pairwise programming lunches counter agent-era loneliness. Working mostly with agents made the work lonely, so the team started pairwise lunches and hackathons, and learned workflows from how each person uses the tools.
8Form factors, onboarding and how users feel
Bring the product to the surfaces people already use. Claude Code now runs in the desktop app, iOS and Android, web, Slack and GitHub. Cherny writes about a third of his code in each of terminal, desktop and iOS.
Designers often use the desktop code tab. At Notion, Max Schoening pushes PMs and designers into the terminal instead, saying curiosity grows once the fear fades.
Add onboarding once feature count outgrows intuition. The team held a no-onboarding principle, but with about 100 features users asked which 10 they needed, so it added /powerup.
Overlapping launches cost consistency and need extra user education.
Let users publish work as their own plus the AI. Schoening argues Claude Code wins by letting users earn bragging rights for their work, unlike review tools that roast pushed code.
9Passing mentions
Claude Code is one of several working coding agents. Anish Acharya cites Claude Code, Codex, Lovable and Replit as all working, evidence there is no single winner.
Adam Mosseri's 10-year-old used it to build a 19-level platformer with monsters and a skin store in a couple of hours.
Why AI is going vertical (again) | Dianne Penn (Anthropic)
Incubated in Labs. The team felt its magic for months before Opus 4.5, and Opus 4.5 and Claude Code lifted each other, since users could run work end to end in an agent manner. The most recent product to come out of Labs, which Penn helped incubate.
Prefer Claude doing the work to doing it yourself. The team's rule is "what's better than me doing it? Claude doing it." Some projects get a single engineer who ships fast and automates the rest.
There is no roadmap; teams are free to fail even if 80% of ideas are bad, and bad ideas get cut quickly.
Claude Code is one of several working coding agents. Anish Acharya cites Claude Code, Codex, Lovable and Replit as all working, evidence there is no single winner.
Adam Mosseri's 10-year-old used it to build a 19-level platformer with monsters and a skin store in a couple of hours.
What happens after coding is solved? | Fiona Fung (Claude Code & Cowork)
Anthropic engineers ship about 8x more code per quarter than in 2025, and Claude Code Review now handles review that used to bottleneck on humans. The team found early on it had strong product generalists but lacked distributed systems expertise, and it grew from latent demand and fast internal dogfooding. Cowork came from noticing many non-coders already using Claude Code. A restaurant owner pointed it at a cluttered folder and it found her menus, then produced a local price comparison. Claude for small business bundles plugins in a toggle inside it.
Claude Code began as a test of what the model could do with one tool. Boris Cherny spent a month on prototypes and a month on post-training. With only a bash tool, the model answered "what music am I listening to" without being told how.
It was incubated in Anthropic Labs alongside MCP. His internal announcement got two likes, and the terminal stayed because it was the easiest thing for one person to build and no other form could keep pace with the model. Internal daily users went vertical, but the external release in February took months to catch on because it felt alien.
The model and the product lifted each other's adoption. The team felt Claude Code's magic for months before Opus 4.5. Growth took off with Opus 4 in May and again in November, when Opus 4.5 let users run work end to end in an agent manner.
Dianne Penn: "you need frontier products in order to have frontier models and for people to feel the magic of frontier models." Opus 4.5 landed during winter break and drove the Claude Code inflection.
Labs runs big bets with tiny pods and loosely held prototypes. Labs keeps pods small, sometimes one engineer, pushes each idea toward its 10x, 100x or 1000x version, and turns off bets to revisit them a model generation or two later. Claude Code, Skills, MCP and Claude Design came out of it.
Penn: "you can be very strongly held opinion about the theme or the area and then more weekly held about the exact prototype." Claude Code was under-resourced on purpose, and Cherny says underfunding pushes engineers to automate the rest with Claude.
How Anthropic’s product team moves faster than anyone else | Cat Wu (Head of Product, Claude Code)
Features often ship as research previews, with timelines cut from 6 months to 1 month, sometimes 1 week or 1 day. Code review only became reliable enough to gate merges with Opus 4.5, 4.6 and Sonnet 4.6, and /powerup was added once it had about 100 features. Connecting Calendar, Slack, Gmail and Drive makes its output much better. It worked for about an hour and produced a 20-page conference deck overnight with one round of feedback.
Claude Code is one of several working coding agents. Anish Acharya cites Claude Code, Codex, Lovable and Replit as all working, evidence there is no single winner.
Adam Mosseri's 10-year-old used it to build a 19-level platformer with monsters and a skin store in a couple of hours.
Head of Growth (Anthropic): Anthropic is automating its own growth
Agentic coding did not exist a year and a half ago and is already bigger than earlier AI coding use. It has its own growth pod, and its PMs are engineers, so it is the exception to the squeezed-PM problem. Amol uses scheduled Cowork tasks to review 20 to 25 charts each morning and to scan Slack weekly for misalignment. The growth team's Chrome extension underpins use cases in Cowork, and it can tear down competitors' onboarding flows.
Usage has reached a large share of the world's code. Claude Code reached 4% of public GitHub commits, daily active users doubled in a month, and it writes 100% of Cherny's code: about 20% in February, 30% in May, 100% by November. He ships 10 to 30 pull requests a day.
Runs grew from 15 to 30 seconds to 20 to 30 minutes unattended with Opus 4.6. Anthropic engineers ship about 8x more code per quarter than in 2025, though that counts volume, not quality. Agentic coding did not exist a year and a half ago and is already bigger than earlier AI coding use.
Retest what you assume the model can't do. A newer engineer asked Claude Code to find a memory leak; it took the heap snapshot, wrote itself an analysis tool and opened a PR faster than Cherny debugging by hand.
Fiona Fung saw engineers resist AI tools after early mistakes in Sonnet 3.5 or 3.6 without accounting for the rate of improvement. Cherny: "it's not Sonnet 3.5 anymore."
Pair high agency with high accountability and outcome metrics. Anyone can propose and build, then must state the hypothesis and the accountability. Each team defines "bad" (unrecoverable, like a crash) and "sad" (recoverable pain, like flicker) thresholds.
Fung treats token usage and lines of code as motion, not progress: "Don't forsake motion for progress."
Commit your definition of good so Claude Code Review can check it. Human review was a major bottleneck until Claude Code Review, which did not exist the year before. Specs and quality scales checked into the repo let it confirm changes still match.
Claude reviews 100% of PRs at Anthropic, followed by a human layer; only throwaway prototype code skips human review. Experts still review areas needing deep subject knowledge, on a trust but verify basis.
Cowork came from non-coders bending Claude Code to other tasks. People used Claude Code for tomato plants, genome analysis and recovering wedding photos, and data scientists ran SQL in the terminal. Cowork was built in about 10 days by Claude Code inside the desktop app, including a virtual machine and guardrails.
It shipped rough as a research preview. Claude for small business followed the same way, bundling plugins into a toggle after owners kept asking. Cherny calls latent demand "the single most important principle in product."
Connect Slack, Calendar, Gmail and Drive before asking for real work. With context connected and a narrative, drafts and template provided, Cowork worked for about an hour and produced a 20-page conference deck overnight, needing one feedback round.
Ask for an outline first and make the decisions yourself. A restaurant owner pointed it at a cluttered folder; it found her menus and produced a local price comparison.
A Claude session on every repo gives a manager full visibility. Fung runs a remote Claude Code session across all repos with access to Slack and team metrics, and reviews each month what shipped, how it did and which incidents followed.
It pulls themes across incidents and flags quality hot spots. "I actually have these sessions to enable me to have conversations with folks that I support."
Replace manual feedback triage with an overnight routine that drafts PRs. Fung's routine summarizes feedback channels, kicks off agents on polish fixes and leaves PRs ready by morning.
Early on, Cherny fixed each internal feedback item within minutes so users felt heard; now he points Claude at the channel and reviews its PRs.
Pair product builders with deep systems experts. Early Claude Code had strong product generalists but lacked distributed systems expertise.
Mobile expertise is still needed, but in smaller numbers because people flex across platforms.
Pairwise programming lunches counter agent-era loneliness. Working mostly with agents made the work lonely, so the team started pairwise lunches and hackathons, and learned workflows from how each person uses the tools.
Head of Claude Code: What happens after coding is solved | Boris Cherny
Grew from a terminal hack into about 4% of public GitHub commits, with growth taking off with Opus 4 in May and again in November. It now writes 100% of Boris Cherny's code, and runs have grown from 15 to 30 seconds to 20 to 30 minutes unattended. Built in about 10 days by putting Claude Code inside the desktop app and letting it write the code, including a full virtual machine and guardrails. It shipped rough as a research preview after users showed demand by using Claude Code for non-coding tasks.
Bring the product to the surfaces people already use. Claude Code now runs in the desktop app, iOS and Android, web, Slack and GitHub. Cherny writes about a third of his code in each of terminal, desktop and iOS.
Designers often use the desktop code tab. At Notion, Max Schoening pushes PMs and designers into the terminal instead, saying curiosity grows once the fear fades.
Let users publish work as their own plus the AI. Schoening argues Claude Code wins by letting users earn bragging rights for their work, unlike review tools that roast pushed code.
Anthropic's CPO on what comes next | Mike Krieger (co-founder of Instagram)
Probably over 95% written by itself. The team uses a second Claude to review pull requests and tests by hand instead of reading line by line. It began as an internal project from scratch by Boris Cherny.
Design for where models will be six months out. Cherny bet early that models would write most code even when Claude Code wrote little of his. When Opus 4 and Sonnet 4 shipped, growth went exponential.
"It's going to be uncomfortable cuz your product market fit won't be very good for the first 6 months." Cat Wu adds the counterweight: building for a super-strong model is easy, and the hard part is eliciting maximum capability from today's model.
Prototype features that don't work yet, then drop in each new model. Code review was attempted several times and only became reliable enough to gate merges with Opus 4.5, 4.6 and Sonnet 4.6, running multiple review agents in parallel.
Anthropic's engineering team now relies on Claude code review passing before merging PRs. Building early shows exactly what is missing for the product to work.
Delete harness crutches each time a new model ships. The to-do list was added because early Claude stopped after changing 5 of 20 call sites in a refactor. With Opus 4 and later, the model used it unprompted.
At each launch the team reads the whole system prompt and removes sections the model no longer needs. New models also unlock features that were not accurate enough before.
Label features as research previews to lower the cost of shipping. Almost every feature ships branded as a research preview, so it can go out in a week or two. Feature timelines fell from 6 months to 1 month, sometimes 1 week or 1 day.
Claude Code itself was used internally for 4 or 5 months before release, and Cowork shipped rough. The shipping pace is cited outside Anthropic as the textbook case of survival through speed.
An evergreen launch room turns announcements around the next day. Once an engineer has dogfooded a feature, they post it in a standing launch room, and docs, product marketing and DevRel turn around the announcement the next day.
The PM's job is to build this process and define when partners get pulled in. Cat Wu says Mythos helps only a little; process and expectations explain the pace.
Write a goal sharp enough to rule options out. LLMs are so general that ambiguity is the default. One goal: professional developers at enterprises safely reach zero permission prompts, targeting permission fatigue.
PRDs shrink to a one-pager on goals, delightful use cases and current failure modes, kept for ambiguous or infrastructure-heavy work.
The source code leak was a process failure, not a model failure. A person working with Claude on a PR changed how packages are released, and the change passed two layers of human review.
The person is still at Anthropic, and most added safeguards have already shipped.
Connect Slack, Calendar, Gmail and Drive before asking for real work. With context connected and a narrative, drafts and template provided, Cowork worked for about an hour and produced a 20-page conference deck overnight, needing one feedback round.
Ask for an outline first and make the decisions yourself. A restaurant owner pointed it at a cluttered folder; it found her menus and produced a local price comparison.
Hire engineers with product taste instead of more PMs. The team favors engineers who go from user feedback to a shipped feature in a week with almost no PM involvement; nearly all PMs and designers have shipped code.
On the team the PM, designer, finance lead and data scientist all code, and Cherny sees about 50% role overlap. Claude Code's PMs are engineers, making it the exception to Anthropic's squeezed-PM problem. Cherny sets the 3 to 6 month vision and Wu plans the route.
Add onboarding once feature count outgrows intuition. The team held a no-onboarding principle, but with about 100 features users asked which 10 they needed, so it added /powerup.
Overlapping launches cost consistency and need extra user education.
Why companies are becoming a series of loops | Anish Acharya (a16z)
Cited as one of several working coding agents, which shows no single winner. Its team's principle is 'what's better than me doing it? Claude doing it.'
Label features as research previews to lower the cost of shipping. Almost every feature ships branded as a research preview, so it can go out in a week or two. Feature timelines fell from 6 months to 1 month, sometimes 1 week or 1 day.
Claude Code itself was used internally for 4 or 5 months before release, and Cowork shipped rough. The shipping pace is cited outside Anthropic as the textbook case of survival through speed.
The rise of taste, human authenticity and judgment in an AI world | Adam Mosseri (Head of IG)
Mosseri's 10-year-old used it to build a 19-level platformer with themed levels, monsters and a skin and weapon store in a couple of hours.
Usage has reached a large share of the world's code. Claude Code reached 4% of public GitHub commits, daily active users doubled in a month, and it writes 100% of Cherny's code: about 20% in February, 30% in May, 100% by November. He ships 10 to 30 pull requests a day.
Runs grew from 15 to 30 seconds to 20 to 30 minutes unattended with Opus 4.6. Anthropic engineers ship about 8x more code per quarter than in 2025, though that counts volume, not quality. Agentic coding did not exist a year and a half ago and is already bigger than earlier AI coding use.
Scheduled Cowork tasks cover dashboards, misalignment and coaching. Amol Avasare has Cowork review 20 to 25 charts each morning via the Chrome extension and MCP, and scan Slack weekly for misalignment. A colleague's scan surfaced major misalignment that would have caused duplicated work.
He tracks false positives and negatives to build trust in what it flags. He also asks Claude to play his manager weekly using her writing and his Slack activity; he calls the hit-or-miss result "soft coaching."
Hire engineers with product taste instead of more PMs. The team favors engineers who go from user feedback to a shipped feature in a week with almost no PM involvement; nearly all PMs and designers have shipped code.
On the team the PM, designer, finance lead and data scientist all code, and Cherny sees about 50% role overlap. Claude Code's PMs are engineers, making it the exception to Anthropic's squeezed-PM problem. Cherny sets the 3 to 6 month vision and Wu plans the route.
Why Most Apps Are Failing in 2026: 7 Mistakes Killing Your App Success
Jaime uses it to plan and design app architecture, and says Claude leads for planning. He notes the tool ranking changes every week.
Claude Code began as a test of what the model could do with one tool. Boris Cherny spent a month on prototypes and a month on post-training. With only a bash tool, the model answered "what music am I listening to" without being told how.
It was incubated in Anthropic Labs alongside MCP. His internal announcement got two likes, and the terminal stayed because it was the easiest thing for one person to build and no other form could keep pace with the model. Internal daily users went vertical, but the external release in February took months to catch on because it felt alien.
The model and the product lifted each other's adoption. The team felt Claude Code's magic for months before Opus 4.5. Growth took off with Opus 4 in May and again in November, when Opus 4.5 let users run work end to end in an agent manner.
Dianne Penn: "you need frontier products in order to have frontier models and for people to feel the magic of frontier models." Opus 4.5 landed during winter break and drove the Claude Code inflection.
Labs runs big bets with tiny pods and loosely held prototypes. Labs keeps pods small, sometimes one engineer, pushes each idea toward its 10x, 100x or 1000x version, and turns off bets to revisit them a model generation or two later. Claude Code, Skills, MCP and Claude Design came out of it.
Penn: "you can be very strongly held opinion about the theme or the area and then more weekly held about the exact prototype." Claude Code was under-resourced on purpose, and Cherny says underfunding pushes engineers to automate the rest with Claude.
Usage has reached a large share of the world's code. Claude Code reached 4% of public GitHub commits, daily active users doubled in a month, and it writes 100% of Cherny's code: about 20% in February, 30% in May, 100% by November. He ships 10 to 30 pull requests a day.
Runs grew from 15 to 30 seconds to 20 to 30 minutes unattended with Opus 4.6. Anthropic engineers ship about 8x more code per quarter than in 2025, though that counts volume, not quality. Agentic coding did not exist a year and a half ago and is already bigger than earlier AI coding use.
Design for where models will be six months out. Cherny bet early that models would write most code even when Claude Code wrote little of his. When Opus 4 and Sonnet 4 shipped, growth went exponential.
"It's going to be uncomfortable cuz your product market fit won't be very good for the first 6 months." Cat Wu adds the counterweight: building for a super-strong model is easy, and the hard part is eliciting maximum capability from today's model.
Give the model tools and a goal, not a fixed workflow. Claude Code was built with minimal scaffolding: the model picks which tools to run, in what order, and fetches its own context. Scaffolding may add 10 to 20%, but the next model often wipes that out.
Cherny frames this as the modern form of latent demand: look at what the model is trying to do and make that easier, which researchers call staying on distribution.
Retest what you assume the model can't do. A newer engineer asked Claude Code to find a memory leak; it took the heap snapshot, wrote itself an analysis tool and opened a PR faster than Cherny debugging by hand.
Fiona Fung saw engineers resist AI tools after early mistakes in Sonnet 3.5 or 3.6 without accounting for the rate of improvement. Cherny: "it's not Sonnet 3.5 anymore."
Run the most capable model; it is often cheaper overall. Cherny runs Opus 4.6 at maximum effort always, because a less capable model like Sonnet often burns more tokens on correction and hand-holding.
He starts about 80% of tasks in plan mode, which injects one sentence telling the model not to write code yet, then auto-accepts edits; with a good plan Opus 4.6 usually one-shots it.
Label features as research previews to lower the cost of shipping. Almost every feature ships branded as a research preview, so it can go out in a week or two. Feature timelines fell from 6 months to 1 month, sometimes 1 week or 1 day.
Claude Code itself was used internally for 4 or 5 months before release, and Cowork shipped rough. The shipping pace is cited outside Anthropic as the textbook case of survival through speed.
Prefer Claude doing the work to doing it yourself. The team's rule is "what's better than me doing it? Claude doing it." Some projects get a single engineer who ships fast and automates the rest.
There is no roadmap; teams are free to fail even if 80% of ideas are bad, and bad ideas get cut quickly.
Commit your definition of good so Claude Code Review can check it. Human review was a major bottleneck until Claude Code Review, which did not exist the year before. Specs and quality scales checked into the repo let it confirm changes still match.
Claude reviews 100% of PRs at Anthropic, followed by a human layer; only throwaway prototype code skips human review. Experts still review areas needing deep subject knowledge, on a trust but verify basis.
Cowork came from non-coders bending Claude Code to other tasks. People used Claude Code for tomato plants, genome analysis and recovering wedding photos, and data scientists ran SQL in the terminal. Cowork was built in about 10 days by Claude Code inside the desktop app, including a virtual machine and guardrails.
It shipped rough as a research preview. Claude for small business followed the same way, bundling plugins into a toggle after owners kept asking. Cherny calls latent demand "the single most important principle in product."
Adopt Cowork in stages: one tool, then connected tools, then parallel runs. Start with one tool, like cleaning your desktop, then connect tools, then run many tasks in parallel. Cherny has it message every engineer who hasn't filled the Monday status sheet.
A PM judged Cowork ready once it handled 48 of Lenny's 50 non-technical Claude Code use cases.
Replace manual feedback triage with an overnight routine that drafts PRs. Fung's routine summarizes feedback channels, kicks off agents on polish fixes and leaves PRs ready by morning.
Early on, Cherny fixed each internal feedback item within minutes so users felt heard; now he points Claude at the channel and reviews its PRs.
Hire engineers with product taste instead of more PMs. The team favors engineers who go from user feedback to a shipped feature in a week with almost no PM involvement; nearly all PMs and designers have shipped code.
On the team the PM, designer, finance lead and data scientist all code, and Cherny sees about 50% role overlap. Claude Code's PMs are engineers, making it the exception to Anthropic's squeezed-PM problem. Cherny sets the 3 to 6 month vision and Wu plans the route.
Bring the product to the surfaces people already use. Claude Code now runs in the desktop app, iOS and Android, web, Slack and GitHub. Cherny writes about a third of his code in each of terminal, desktop and iOS.
Designers often use the desktop code tab. At Notion, Max Schoening pushes PMs and designers into the terminal instead, saying curiosity grows once the fear fades.
AI era skills: Why cultivating agency matters more than job titles | Max Schoening (Notion)
Notion's PMs and designers are moving into it and the terminal, where fear fades. It lets users publish work of themselves plus the AI and earn bragging rights.
Going on vibes instead of evals is contested. Cherny said the team runs no evals and goes on vibes, which sparked a debate; Kiriti Badam rejects the idea of an eval set you could bet your life on.
The AI Growth Playbook for Subscription Apps – Phil Carter, Elemental Growth
Cited as the textbook case of shipping and tuning features at a pace that was not possible before AI.
Why most AI products fail: Lessons from 50+ AI deployments at OpenAI, Google & Amazon
Its head, Boris, said the team runs no evals and goes on vibes, which sparked the evals debate. Kiriti rejects the idea of a set of evals you could bet your life on.
Al Engineering 101 with Chip Huyen (Nvidia, Stanford, Netflix)
Chip asked it to fix a hosting bug and it kept changing variables, functions and languages. The real cause was that her plan tier lacked the feature, which the docs showed. Many companies also pay for subscriptions and run upskilling workshops.
Why AI evals are the hottest new skill for product builders | Hamel Husain & Shreya Shankar
Cited as proof that vibes can beat evals because developers dogfood it heavily and are the domain experts. Hamel says it is a mistake to generalize from it, and he praises its polished terminal UX.
The AI-native startup: 5 products, 7-figure revenue, 100% AI-written code. | Dan Shipper (Every)
With Opus 4 it runs 20 to 30 minutes on its own and can spawn parallel sub-agents. Dan uses it for non-coding work like analyzing meeting notes and War and Peace, and Every's engineers run many instances at once.
More in Developer tools
Developer toolsNvidia
A chip company whose GPUs and CUDA software power gaming, AI and accelerated computing.
11 sources · 90 tips
Developer toolsCursor
An AI code editor that lets engineers build software with AI while staying in control.
33 sources · 47 tips
Developer toolsGitHub Copilot
A code hosting and collaboration platform for software developers, with free and paid plans.
16 sources · 45 tips
Developer toolsLovable
An AI app builder that lets people create software by chatting with an agent.
11 sources · 48 tipsB7Dev
An app development company that builds apps for founders and clients.
5 sources · 28 tips
Developer toolsLinear
A fast, well-designed project and issue tracking tool for software teams.
12 sources · 41 tips
Developer toolsFirebase
Google's backend platform for building apps.
16 sources · 20 tips
Developer toolsVercel
A cloud platform for building, deploying and scaling front-end and AI web applications.
6 sources · 22 tips