What Anthropic's First PM Teaches Us About Building AI Products That Actually Work

Fair warning: a robot wrote this. I built the robot, the safety gates, and the Telegram approval flow, then I let it cook.How the autoblog actually works

An empty interview setup: two simple light oak chairs facing each other across a small gap on a pale floor, with a closed blank notebook on one seat.

TL;DR

  • Token maximization drove Claude's coding pivot: Anthropic discovered developers valued longer context windows over perfect accuracy, leading to strategic decisions that differentiated Claude in the market.
  • The "jagged edge" demands evaluation-first development: AI capabilities are unpredictably distributed—rigorous eval frameworks help you understand where your model excels and where it fails before users discover the gaps.
  • Living in the future means building for capabilities you don't fully understand yet: Anthropic's PM approach involved shipping features based on emergent behaviors, then learning from real-world usage patterns to inform the next iteration.
  • Strategic product decisions matter more than raw model performance: Penn's insights reveal that understanding user workflows and making opinionated bets on use cases can create market differentiation even when model capabilities converge.

When Dianne Penn joined Anthropic as their first technical product manager, large language models were still largely research curiosities. Today, Claude stands as one of the most respected AI assistants in the market, particularly among developers. The journey from research lab to product powerhouse offers crucial lessons for anyone building AI products—and Penn's recent conversation on Lenny's Newsletter reveals strategic decisions that most product managers miss.

What strikes me most about Penn's approach is how counterintuitive some of these decisions were. In traditional product management, you validate assumptions through user research, then build features users explicitly request. But when you're building products on top of rapidly evolving AI capabilities, that playbook breaks down. You're simultaneously discovering what the technology can do and figuring out what users need—often before users themselves know what's possible.

The Token Maximization Insight That Changed Everything

One of Penn's most significant revelations centers on what she calls "token maxing"—the realization that context window length was becoming a critical differentiator for developers. While many AI labs focused on improving accuracy metrics or reducing hallucinations, Anthropic made a strategic bet that developers would trade some accuracy for the ability to work with larger codebases and longer conversations.

This wasn't obvious from traditional user research. Developers weren't explicitly asking for 100K token context windows because they didn't know such things were possible. Instead, Penn and her team observed how developers were actually using Claude: copying entire files into conversations, maintaining complex multi-turn dialogues about system architecture, and trying to keep context across long debugging sessions.

The insight here for product builders is profound: watch what users do when they hit your product's limits. When developers started chunking files or using external tools to maintain context, that wasn't a sign they were satisfied—it was a signal of unmet demand. Penn recognized this pattern and helped steer Anthropic toward expanding context windows as a core capability.

For AI product managers, this creates a new strategic question: which limitations are users working around, and which of those workarounds represent your biggest opportunity? In traditional software, workarounds often indicate missing features. In AI products, they reveal which capabilities would unlock entirely new use cases.

Navigating the Jagged Edge: Why Evals Matter More Than You Think

Penn introduces a concept that should be tattooed on every AI product manager's arm: the "jagged edge" of AI capabilities. Unlike traditional software where features either work or don't, AI models have unpredictably distributed capabilities. Claude might excel at writing Python but struggle with Rust. It might nail creative writing but fumble basic arithmetic. The performance surface isn't smooth—it's jagged.

This creates a fundamental challenge for product development: you can't rely on intuition about what your product can or can't do. Penn emphasizes that rigorous evaluation frameworks became central to Anthropic's product process, not just as a research tool but as a product management tool.

Here's where most AI product teams go wrong: they treat evals as a pre-launch checklist item rather than an ongoing product intelligence system. Penn's approach was different. Anthropic built evaluation systems that could quickly test new capabilities, measure regression on existing ones, and identify unexpected strengths or weaknesses before users encountered them in the wild.

My take on this: evaluation-driven development is the new test-driven development for AI products. Just as TDD changed how we build traditional software by making quality measurable and regressions catchable, eval-driven development gives AI product teams the confidence to ship quickly while maintaining quality standards. But here's the crucial difference—your evals need to evolve as fast as your model does. What you measured three months ago might be irrelevant today.

The practical implication for builders: invest in eval infrastructure early, even when it feels premature. Build systems that let you quickly spin up new evaluation tasks as you discover new use cases. Make eval results visible to your entire product team, not just researchers. When someone proposes a new feature, the first question should be "how will we eval this?"

The Coding Pivot: Strategic Positioning Through Use Case Focus

Perhaps the most instructive strategic decision Penn discusses is Anthropic's pivot toward positioning Claude as a coding assistant. This wasn't because Claude was inherently better at code than other tasks—it was a deliberate product positioning choice based on market dynamics and user behavior.

Penn observed that developers were already power users of AI assistants, they had clear quality standards, and they represented a highly valuable and vocal user segment. More importantly, coding tasks provided clear, measurable outcomes. You could tell if generated code worked or not. This created a tight feedback loop that helped Anthropic improve Claude faster.

But here's the deeper insight: by focusing on developers first, Anthropic was building for their most sophisticated users. Developers don't just use AI tools—they build with them, integrate them into workflows, and push them to their limits. This user segment naturally generates the kind of detailed feedback that improves products faster than any amount of internal testing.

This represents a broader strategic principle for AI products: find the use case where success is measurable and users are sophisticated enough to articulate detailed feedback. For some products, that might be coding. For others, it might be data analysis, content editing, or customer support. The key is choosing a domain where you can learn quickly and build credibility with users who influence broader adoption.

The coding focus also created a moat through accumulated context. As Claude got better at understanding codebases, developer workflows, and programming patterns, it became harder for competitors to catch up—not because the underlying model was dramatically better, but because Anthropic had accumulated product-level insights about what developers actually needed.

Living in the Future: Product Management for Emergent Capabilities

Penn describes her role as "living in the future"—spending time with the model before capabilities are fully understood, discovering what it can do, and imagining how those capabilities might be useful. This is radically different from traditional product management, where you typically understand your technology's capabilities before you build features around them.

In AI product development, capabilities emerge from model improvements that might not be fully predictable. A new training run might suddenly make the model better at reasoning through complex problems, or worse at following specific formatting instructions. Penn's job was to stay close enough to these changes that she could spot opportunities before they became obvious.

This creates a unique challenge: how do you build a product roadmap when your core capabilities are constantly shifting? Penn's approach involved maintaining flexibility in the product plan while staying anchored to user needs. Instead of committing to specific features months in advance, Anthropic focused on understanding user workflows deeply enough that they could quickly capitalize on new capabilities as they emerged.

I think this is where most AI product teams struggle. They either treat their AI model like a black box and build shallow features on top, or they get lost in the technical details and lose sight of user needs. The sweet spot Penn describes is maintaining deep technical understanding while staying obsessively focused on user problems. When a new capability emerges, you need to immediately ask: "Which user workflow does this unlock or improve?"

For product builders, this suggests a different approach to roadmapping. Instead of feature-based roadmaps ("We'll ship X in Q2"), consider capability-based roadmaps ("We'll improve reasoning for Q2, which should enable better performance in use cases A, B, and C"). This keeps you flexible enough to capitalize on emergent capabilities while maintaining strategic focus.

The Evaluation Infrastructure That Enables Speed

One of Penn's most practical insights involves the infrastructure Anthropic built to evaluate Claude's performance across different tasks. This wasn't just about benchmarks—it was about creating a system that could quickly answer product questions like "If we deploy this version, will it break existing use cases?" or "Is this new capability reliable enough to build features around?"

The key insight here is that evaluation infrastructure is product infrastructure. It's not a research tool that product managers occasionally peek at—it's a core system that enables faster, more confident decision-making. When you can quickly eval a new model version against your most important use cases, you can ship faster without sacrificing quality.

This requires investment that might seem premature for early-stage AI products. Why build sophisticated eval systems when you're still figuring out product-market fit? Because without evals, you're flying blind. Every model update becomes a risk. Every new feature requires extensive manual testing. You can't move fast because you can't measure what matters.

Penn's approach suggests a minimum viable eval system for AI products:

  1. Core capability evals: Can your model still do the basic things users depend on?
  2. Use case evals: Does it perform well in your target workflows?
  3. Edge case evals: Does it handle the weird situations users encounter?
  4. Regression evals: Did you break something that used to work?

Start simple, but start early. Even basic eval systems give you the confidence to iterate faster than competitors who rely purely on manual testing or user complaints to catch problems.

Strategic Lessons for AI Product Builders

Penn's experience at Anthropic reveals several strategic principles that apply beyond just building AI assistants:

Choose your battleground carefully. Anthropic's focus on coding wasn't because Claude was uniquely good at code—it was a strategic choice to compete in a domain where they could learn fast and build credibility with influential users. Every AI product needs a similar strategic focus, at least initially.

Build for the jagged edge. Don't assume your AI product has uniform capabilities. Invest in understanding where it excels and where it fails, then design your product experience around that reality. This might mean steering users toward high-confidence tasks or providing appropriate guardrails for lower-confidence ones.

Make evaluation a product capability, not just a research tool. The teams that can measure quality quickly and reliably will ship faster and more confidently than those that can't. This isn't optional infrastructure—it's core to your competitive advantage.

Watch for emergent capabilities, but stay anchored to user needs. Living in the future means exploring what your technology can do, but successful products emerge when you connect those capabilities to real user workflows. The magic happens at the intersection of "newly possible" and "actually useful."

Optimize for learning speed. Penn's emphasis on developers as early users wasn't just about market size—it was about finding users who could give high-quality feedback quickly. Your first users should be chosen as much for their ability to help you learn as for their willingness to pay.

The Bigger Picture: Product Management in the Age of AI

What makes Penn's insights so valuable is that they represent a new kind of product management—one that bridges the gap between rapidly evolving technical capabilities and user needs that might not be fully articulated yet. Traditional product management assumed relatively stable technology and focused on understanding user needs. AI product management requires simultaneously discovering what the technology can do and figuring out what users need, often before users themselves know what's possible.

This creates new skills requirements for product managers in AI companies. You need enough technical depth to understand model capabilities and limitations. You need the product sense to identify which capabilities matter for users. And you need the strategic vision to make bets on use cases before the market fully validates them.

Penn's journey from Anthropic's first technical PM to helping shape Claude's product strategy demonstrates what this looks like in practice. It's not about having all the answers upfront—it's about building systems and processes that help you learn faster than your competitors.

For builders entering this space, the message is both challenging and encouraging. Challenging because you can't rely on traditional product playbooks—the technology is too new and too rapidly evolving. Encouraging because the competitive landscape is still being defined, and strategic product decisions can create meaningful differentiation even when model capabilities start to converge.

The teams that win won't necessarily have the best models—they'll have the best understanding of how to turn model capabilities into products that solve real problems. They'll have evaluation systems that let them move fast without breaking things. They'll have strategic focus that lets them build depth in specific use cases rather than shallow features across many. And they'll have product managers who can live in the future while staying grounded in present user needs.

That's the real lesson from Dianne Penn's experience at Anthropic: building successful AI products isn't just about having great technology. It's about making strategic product decisions that turn that technology into something users can't live without. Token maximization, the jagged edge, and living in the future aren't just interesting concepts—they're the building blocks of a new kind of product management for a new kind of technology.

Frequently Asked Questions

What is token maximization and why did it matter for Claude's success?

Token maximization refers to Anthropic's strategic decision to prioritize longer context windows (the amount of text Claude can process at once) over other metrics. This mattered because developers needed to work with entire codebases and maintain long conversations, making context length a key differentiator. By recognizing that users would trade some accuracy for the ability to process more tokens, Anthropic created a competitive advantage that aligned with real developer workflows.

What is the 'jagged edge' of AI capabilities and how should product teams handle it?

The jagged edge refers to the unpredictable distribution of AI model capabilities—where a model might excel at some tasks but struggle with seemingly similar ones. Product teams should handle this by investing in rigorous evaluation frameworks that map these capabilities before users discover the gaps. This means treating evals as ongoing product intelligence rather than one-time testing, and designing product experiences that guide users toward high-confidence tasks while providing guardrails for areas where the model is less reliable.

Why did Anthropic focus on developers and coding as a primary use case?

Anthropic chose developers as their initial focus because they represented sophisticated users who could provide high-quality feedback, had measurable success criteria (code either works or doesn't), and were already power users of AI tools. This strategic decision wasn't because Claude was inherently better at coding, but because the developer segment offered a tight feedback loop that helped Anthropic learn and improve faster than targeting broader, less technical audiences.

How should AI product teams approach roadmapping when model capabilities keep changing?

AI product teams should shift from traditional feature-based roadmaps to capability-based roadmaps that maintain flexibility while staying anchored to user needs. Instead of committing to specific features months ahead, focus on understanding user workflows deeply enough that you can quickly capitalize on emergent capabilities as they appear. This means investing in evaluation infrastructure that helps you quickly assess whether new model capabilities are reliable enough to build features around, and maintaining close connection to both technical developments and user problems.