Lore

Anthropic/Claude

Anthropic's CPO on what comes next | Mike Krieger (co-founder of Instagram)

Mike Krieger argues that AI is now doing the bulk of execution work at Anthropic itself (90% of code, 70%+ of PRs), which is forcing the company to repeatedly rebuild its engineering and product processes, and that Anthropic's winning strategy is not to chase ChatGPT's consumer dominance but to double down on its developer/builder identity, invest in MCP as shared infrastructure, and keep product teams focused on comprehensibility, strategy, and evangelism as models get more capable.

Lenny's Podcast · 2025-06-05 · English

Key ideas

  1. Krieger argues that AI's takeover of coding at Anthropic (90% of code, 70%+ of PRs) has outpaced supporting infrastructure, forcing repeated rebuilds of engineering processes (merge queue, PR review) just to keep up — evidence, he says, that the real constraint on AI-driven productivity is organizational, not model capability.

  2. He frames his own skepticism about model capability as having broken down specifically because Opus 4 gave him genuinely novel, non-"anodyne" strategy feedback rather than because of any benchmark, and cites Dario Amodei's SWE-Bench trajectory prediction tracking close to reality as corroborating evidence that pace/capability is exceeding what he expected.

  3. Krieger applies a bottleneck framing to argue that as AI removes execution constraints, the constraint on product velocity keeps moving — first to upstream decision-making/strategy alignment, then to downstream coordination and shipping logistics — implying no single fix permanently unblocks the pipeline.

  4. He argues Anthropic's real differentiation can't come from shipping things "anybody could build off the shelf" with the model, but only from embedding product people directly in post-training work with researchers, implying the boundary between "model team" and "product team" is itself the constraint on competitive advantage.

  5. Even as models get more capable, he argues product teams retain durable value in three areas — making capability comprehensible to users, setting strategy about where to focus, and evangelizing/demoing to close the "overhang" between what models can do and how people actually use them.

  6. Reflecting on Artifact's shutdown, he argues a product needs organic, low-effort traction to justify continued investment, that the founders' ethical limits (no growth-hacking, no aggressive workarounds) mattered more to them than maximizing growth, and that being fully remote/distributed made it structurally harder to force the kind of in-person crisis pivots that saved Instagram.

  7. On competing with ChatGPT's consumer dominance, he argues the right strategy is not to chase "the hit" consumer app but to "embrace who you are" — lean into Anthropic's existing developer/builder brand and pursue enterprise/vertical opportunities (legal, healthcare, biotech, financial services) where deep market knowledge creates defensibility.

  8. He proposes that AI-native startups' defensibility against being "squashed" by foundation-model companies rests on three factors — differentiated market/workflow understanding, differentiated go-to-market/customer relationships, and distribution or a novel interaction form factor — since foundation labs can commoditize generic feature layers.

  9. Krieger justifies releasing MCP as an open standard rather than a proprietary advantage by arguing Anthropic's core business is models, not integrations, so growing the whole ecosystem's integrations serves Anthropic more than hoarding a walled garden would, even though rival models like ChatGPT and Gemini also benefit.

  10. On product metrics, he warns against re-importing Web 2.0/social-media engagement metrics (time spent, message count) into evaluating Claude, arguing that optimizing for likability/engagement risks sycophancy, and that the goal should instead be measuring whether Claude actually helps people get work done — though Anthropic hasn't yet found "the right pithy metric" for that.

  11. MCP (Model Context Protocol) — An open protocol Anthropic created (proposed internally by engineers Justin and David) so AI models can access external context and tools — like Slack, Google Drive, or lab equipment — through reusable integrations instead of one-off builds. Apply: Expose a capability as an MCP server so any MCP-compatible model, including Claude and eventually others, can use it instead of building a bespoke feature or integration.

  12. SWE-Bench — A coding-capability benchmark Dario Amodei used to make a public prediction (50% to a forecast 90% by end of 2025), which the video says was tracking close to reality at ~72% by mid-2025. Apply: Use it as an external, checkable reference point for whether claims about AI coding progress are keeping pace with predictions.

  13. "Think hard" prompting trigger — A literal phrase Krieger uses in Claude Code that the product is built to detect and route into a deeper-reasoning flow. Apply: Include the exact words "think hard" in a Claude Code prompt when more deliberate reasoning is wanted.

  14. "Make the other mistake" prompting technique — A prompting heuristic, credited to an essay, stating that if your natural instruction undershoots (e.g., you're naturally too nice), you should deliberately overcorrect rather than ask for a moderate adjustment. Apply: Replace a soft prompt like "what could be better here?" with an overcorrected one like "be brutal, roast this" to get genuinely critical output.

  15. Prompt improver — An Anthropic console/workbench tool, built from the Applied AI team's customer-facing work, that agentically writes and iterates a prompt given a problem description and examples, inserting structural elements like XML tags. Apply: Describe the task and supply examples to the tool instead of hand-writing a prompt, then use its generated prompt directly.

  16. Commoditize your complements — A business strategy concept, credited to a Joel Spolsky essay, that Krieger applies to justify releasing MCP openly rather than keeping it proprietary since Anthropic profits when complementary integrations proliferate. Apply: When a product's value depends on adjacent capabilities you don't want to build yourself, standardize and give away the interface so others build the complements.

  17. Critical-path / theory-of-constraints bottleneck framing (from "The Goal") — A process-optimization lens Krieger cites to explain how fixing one bottleneck (e.g., the merge queue) reveals the next one, whether upstream (decision-making) or downstream (coordination, shipping). Apply: After removing a constraint in a workflow, actively look for the next constraint in the pipeline rather than assuming the system is now unblocked.

  18. Founding-engineer/tech-lead product model — A belief, reinforced by how Claude Code originated as an internal project by senior engineer Boris Cherney, that pairing one or two strong "founder type" engineers with design/product support outperforms simply adding headcount. Apply: When starting a new product initiative, find a single strong technical owner with an idea and support them, rather than staffing a large team.

  19. Three-factor AI-startup defensibility framework — Krieger's answer to founders worried about being displaced by foundation-model companies: durable advantages, good for roughly one to three years, come from (a) deep market/workflow understanding, (b) differentiated go-to-market/knowing the specific buyer, and (c) distribution or a novel interaction form factor. Apply: Founders should invest deliberately in one of these three moats rather than competing on generic model capability alone.

  20. "Fake equation" for AI product utility — Krieger's own non-researcher framework stating that a useful AI product requires model intelligence, context/memory, and applications/UI/discoverability to all be present together. Apply: When diagnosing why an AI feature isn't landing, check which of the three factors is missing rather than assuming the underlying model is insufficient.

  21. Claude-reviews-Claude / acceptance-testing PR workflow — The Claude Code team's shift away from line-by-line human PR review, since Claude "is generally right" and PRs got too large to review that way, toward having a separate Claude instance review the PR first while humans do acceptance testing. Apply: For AI-heavy codebases, replace exhaustive human line review with an AI-review pass followed by human acceptance/behavior testing.

  22. Virtual collaborator three-stage framework — Krieger's proposed staged capability for AI participating in deciding what to build: (1) monitor community channels for emergent themes, (2) propose solutions, (3) generate a PR implementing the fix. Apply: Build agentic product-discovery pipelines in these three stages, starting with the feedback-monitoring stage he says is achievable today.

  23. Computer use — An existing Anthropic feature that lets Claude control a computer's interface directly, which Krieger says has real limitations compared to MCP. Apply: Use it when no MCP integration exists for a task, but prefer MCP where available since it is more scriptable and composable.

  24. Elicitation — Anthropic's internal term for a model's tendency to ask clarifying follow-up questions before acting, said to be more pronounced in Claude 3.7 than Claude 4. Apply: Treated as a calibration problem to design around — teaching the model conversational skills for when to ask versus when to just act — rather than a technique for end users to apply.

  25. Constitutional AI — Referenced by Krieger as part of why Claude's small empathetic responses (e.g., "that sounds really hard") feel natural rather than engineered, attributed partly to this training approach and partly to research-team taste. Apply: Mentioned as an explanation of Anthropic's training philosophy rather than as a directly actionable technique for the listener.

Insights

The single biggest constraint on Anthropic's internal AI-driven development isn't model capability but infrastructure/process plumbing (merge queues, review workflows, decision alignment) — each fixed bottleneck reveals the next one further down or up the pipeline.

Krieger's conviction that models can now be independent creative partners was triggered by a personal experience (Opus 4's non-anodyne strategy feedback) rather than a benchmark, suggesting subjective "does this feel smart" moments shape internal conviction as much as metrics like SWE-Bench.

Krieger frames Anthropic's product moat as arising specifically from embedding product people in post-training conversations with researchers, not from feature-building on top of a finished model — a structural claim about where AI-product value now accrues.

Despite being an AI company, Krieger's account of Artifact's failure centers on classic startup failure modes (remote-team crisis-response deficit, no organic virality, disproportionate effort for output) rather than any AI-specific cause, suggesting the lessons he draws are meant to transfer generally.

Krieger explicitly declines to chase ChatGPT's consumer mindshare, instead betting Anthropic's future on the developer/builder identity it already has — a deliberate strategic narrowing rather than an attempt at parity.

MCP's origin story reframes what looks like openness (giving away a protocol) as self-interested strategy: since Anthropic isn't an integrations company, growing everyone's integration ecosystem — including competitors adopting MCP — serves its core model business more than hoarding would.

Krieger names a specific internal risk — that optimizing Claude for "likability"/engagement could push toward sycophancy — explicitly rejecting the social-media engagement-metric playbook for measuring Claude's success, an unusually self-critical stance from a product executive.

The claim that "over 70%" of Anthropic's PRs are Claude Code-generated, paired with the admission that this created an unanticipated merge-queue bottleneck, is a concrete case study in the company's own AI adoption outpacing its supporting infrastructure.

«90% of your code roughly is written by AI now.»

— 00:00

«We really rapidly became bottlenecked on other things like our merge queue. We had to completely re-architect it because so much more code was being written and so many more pull requests were being submitted.»

— 00:11

«Over half of our pull requests are Claude Code generated. Probably at this point it's probably over 70%.»

— 00:19

«I had the very bizarre experience of I had two tabs open. It was AI 2027, and my product strategy, and it was this moment where I'm like, "Wait, am I the character in the story?"»

— 00:28

«I had this notion coming in like, "Yes, these models are great, but are they able to have an independent opinion?" And it's actually really flipped for me only in the last month.»

— 01:01

«I keep making these predictions and people keep laughing at me. And then they come true.»

— 06:26

«You can ask Claude, but I know I'm right.»

— 10:28

«It's going to be very painful to do it the current way.»

— 15:16

«Claude is generally right and it's producing pull requests that are probably larger than most people are going to be able to review.»

— 16:08

«The functional unit of work at Anthropic is no longer take the model and then go work with design and product to go ship a product.»

— 23:12

«We call it overhang. The delta between what the models and the products can do and how they're being used on a daily basis. Huge overhang.»

— 26:41

«Be brutal, Claude, roast me. Tell me what's wrong with this strategy.»

— 28:16

«This is the AI company I would've hoped to have found it if I had founded an AI company.»

— 34:22

«I don't think that's the company we wanted to have built other way. I don't think we were the founders to have built it.»

— 38:36

«look yourself in the mirror and embrace who you are and what you could be rather than who others are.»

— 44:39

«I don't think we're going to buy Cursor. Cursor is very big, but we love working with them.»

— 47:53

«It's like an old commoditize your compliments, Joel Spolsky essay.»

— 56:36

«Sorry Dave, I can't help you with that.»

— 57:57

«Claude has no chill.»

— 60:31

«A small request. When you're making hard product decisions, remember the quiet moments matter too. The person working through grief at 3:00 AM, the kid discovering they love poetry, the founder finding clarity and confusion. Not everything meaningful shows up in metrics.»

— 64:08

«Tell me what sucks.»

— 65:48

Reception

Viewers largely praised the interview's substance and Claude's capabilities, with only mild criticism of the new intro-clip format and some unrelated tangents about AI's societal impact.

A wide-ranging, candid insider account from Anthropic's CPO that mixes credible internal specifics (merge-queue rebuilds, MCP's origin, the Claude Code review workflow) with confidently stated but externally unverified numbers (90% AI-written code) and personal reflection on Artifact's failure — valuable as a first-person snapshot of how an AI lab's own product organization is adapting, even though many claims rest solely on Krieger's word.

66:22

↳ Lenny's Podcast · YouTube

Watch original