interview_prep_doc / art_T3pKoYa_4WA

Content

{
  "markdown": "# Anthropic Interview Prep\n## Questions They May Ask You \u2014 Claude Code PM Interview\n\n---\n\n## 1. Interviewer Summary\n\nBecause this is a general Anthropic interview loop rather than a single named interviewer, expect to face a rotating panel of senior technical leaders, researchers, and PMs from the Claude Code and developer products teams. Anthropic's known loop shape \u2014 an AI engineering screen followed by four onsites capped with a research presentation \u2014 means you will encounter interviewers who span the full spectrum from deep ML research to product strategy.\n\nThe through-line across all of them is **empirical rigor**. Anthropic is not a typical PM-culture company. Interviewers will probe whether your technical claims are real, whether your product instincts are grounded in data, and whether your safety thinking is genuine rather than performative. Surface-level PM frameworks will not pass the bar here.\n\n**What this panel collectively cares about:**\n- Is Felix a genuine technical builder who happens to do PM \u2014 or a PM who talks technical?\n- Can he translate frontier model capabilities into practical developer experiences, not just describe the vision?\n- Does he have a specific, informed point of view on Claude Code's current gaps \u2014 grounded in hands-on use?\n- Is his AI safety alignment real, or is it interview polish?\n- Can he operate effectively inside a structured, research-driven organization after 18 months of founder autonomy?\n\nYour background is unusually strong for this role: NeurIPS-published researcher, 12-algorithm RL workbench, production multi-agent orchestration framework, and Staff PM at Intuit scale. The primary job in the room is to make that depth *visible* \u2014 not just list it, but demonstrate it through the specificity and honesty of your answers.\n\n---\n\n## 2. Questions They May Ask\n\n---\n\n### A. R\u00e9sum\u00e9 Deep Dive\n\n**Question 1.** **Walk me through the ICE platform at Intuit \u2014 specifically, what was the hardest technical decision you made as PM, and how did you make it? What data did you use, and what did you get wrong?**\n\n**Why they may ask:** ICE is your highest-scale platform story (675M+ engagements, 50K TPS, rSocket migration). A technical interviewer will stress-test whether you drove the architecture or just reported on it. The rSocket migration to support 1.5M concurrent connections is a specific, verifiable claim they will probe deeply.\n\n**How to prepare:** Build a crisp narrative with four beats: (1) what the bottleneck was *before* rSocket \u2014 why the existing stack couldn't scale past 6K TPS; (2) why rSocket over alternatives like WebSocket or gRPC \u2014 go to the protocol-level tradeoffs, not just \"it was faster\"; (3) what broke during the migration and how you diagnosed it; (4) what you'd do differently. Be ready to explain sub-25ms TP99 in terms of the specific levers you pulled \u2014 not just the outcome metric.\n\n---\n\n**Question 2.** **Your RL workbench benchmarks GRPO, DPO, PPO, and nine other algorithms across TRL, VeRL, OpenRLHF, and NeMo RL. What did you actually learn from running these benchmarks \u2014 what surprised you about the convergence or throughput differences across frameworks?**\n\n**Why they may ask:** This is a direct credibility probe. Anthropic's research team works on RLHF and RLAIF and will immediately know if you're parroting framework names versus having run real experiments. The question is not about the architecture of the workbench \u2014 it's about your empirical findings.\n\n**How to prepare:** Prepare two or three concrete, specific findings with numbers: e.g., GRPO vs. DPO convergence behavior on GSM8K at a given step count, memory footprint differences across frameworks on Apple Silicon MPS vs. CUDA, or a surprising failure mode (a run that diverged when theory said it shouldn't). Have throughput figures, memory figures, and steps-to-convergence ready. If you don't have a surprising finding, that itself is a finding worth discussing honestly.\n\n---\n\n**Question 3.** **You built OpenClaw as a multi-agent orchestration framework with a gateway protocol and subagent delegation. How does your design compare to how Claude Code handles agentic task decomposition today, and what would you change about Claude Code's approach based on what you learned building OpenClaw?**\n\n**Why they may ask:** This is the intersection of your hands-on agentic engineering and the Claude Code PM role. It tests whether you've used Claude Code deeply enough to have a real point of view, and whether OpenClaw gives you genuine product insight versus r\u00e9sum\u00e9 decoration.\n\n**How to prepare:** Use Claude Code extensively before every interview round \u2014 build something real with it. Map OpenClaw's gateway/subagent model to Claude Code's tool-use and subagent patterns: where do they diverge, and why? Prepare one specific, grounded critique or improvement idea. \"Claude Code's subagent delegation doesn't yet have a clean mechanism for X, and here's how I'd address it based on what I learned building the gateway protocol in OpenClaw\" is the level of specificity that lands.\n\n---\n\n**Question 4.** **Your NeurIPS 2014 paper was on neural networks for protein secondary structure prediction. How has your thinking about neural network design evolved from that work to what you're building now \u2014 and how does that arc inform how you think about evaluating model capabilities for Claude Code?**\n\n**Why they may ask:** The NeurIPS paper gives you peer-level credibility with Anthropic researchers, but they'll want to see intellectual continuity and genuine evolution of thinking \u2014 not just a credential on a r\u00e9sum\u00e9. Your aeval platform and AutoEval work are the natural bridge.\n\n**How to prepare:** Prepare a narrative arc that connects the dots honestly: 2004 hand-coded BPTT in C++ \u2192 NeurIPS 2014 \u2192 aeval's statistical rigor (bootstrap CIs, Welch's t-test, Cohen's d, saturation detection) \u2192 what you'd apply to evaluating Claude Code's agentic task completion. The throughline is \"I've always cared about whether the model is actually doing what we think it's doing \u2014 and measurement is the hard part.\" Show that your eval thinking is principled and empirical, not ad hoc.\n\n---\n\n### B. Technical Domain\n\n**Question 5.** **Claude Code is increasingly used for multi-step agentic tasks \u2014 writing code, running tests, reading errors, iterating. What do you think the right evaluation framework looks like for agentic coding quality, and how would you instrument Claude Code to measure it?**\n\n**Why they may ask:** This is the core technical PM question for the role. Your aeval platform (5 eval types, adversarial safety testing, CI/CD integration) and AutoEval work (reducing eval cycles from 72 hours to 4 minutes) make this directly relevant. Anthropic wants to see whether you can translate eval engineering experience into product metrics.\n\n**How to prepare:** Design a concrete framework on paper before the interview. Candidate metrics: task completion rate, edit distance from gold solution, test pass rate, number of tool calls per task, error recovery rate, user override frequency. Reference your aeval architecture (FastAPI orchestrator, TimescaleDB, Redis job queue) and explain how you'd adapt it for Claude Code's agentic loop \u2014 specifically, how you'd handle the fact that there's often no single \"gold\" solution for a real-world coding task.\n\n---\n\n**Question 6.** **The JD mentions \"building an ecosystem around the CLI so developers can easily share best practices.\" What does that ecosystem actually look like in practice \u2014 what are the primitives, the sharing mechanisms, and how do you prevent it from becoming a graveyard of unused snippets?**\n\n**Why they may ask:** This is a product design plus technical architecture question specific to Claude Code. Your DevPortal work at Intuit (GitOps config, ICE Playground) and SDK scaffolding experience are directly relevant. Anthropic wants to see whether you can think beyond the CLI to ecosystem dynamics \u2014 specifically, what makes shared tooling actually get adopted.\n\n**How to prepare:** Think through the primitives concretely: CLAUDE.md as a sharing unit, community prompt libraries, slash command registries, MCP server marketplaces. Then draw on your Intuit DevPortal experience \u2014 what made developers actually adopt shared tooling versus ignore it? The graveyard problem is real and worth addressing directly: curation, discoverability, and social proof (usage counts, author reputation) are the levers. Prepare a concrete ecosystem proposal with three primitives and their adoption mechanics.\n\n---\n\n**Question 7.** **Anthropic recently acquired Stainless, which builds SDK generation tooling. If you were PM for the Claude Code + Stainless integration, what's the first thing you'd ship and why?**\n\n**Why they may ask:** This tests whether you've done your homework on Anthropic's recent moves and can connect the Stainless acquisition to Claude Code's developer platform strategy. Your Intuit SDK Starter Kit experience (Java/Python scaffolding, Gradle/Maven, CI/CD integration) is the most directly relevant background in your r\u00e9sum\u00e9.\n\n**How to prepare:** Research Stainless's product before the interview \u2014 it auto-generates idiomatic SDKs from OpenAPI specs. Think through the integration surface: Could Claude Code use Stainless to auto-generate client libraries for MCP servers a developer is building? Could Stainless tooling be embedded in Claude Code's agentic workflow so that when Claude Code integrates a new API, it automatically generates a typed client? Have a specific, scoped first ship with a clear success metric \u2014 not a three-year vision.\n\n---\n\n**Question 8.** **You implemented 12 RL algorithms in your workbench. From a product perspective, how do you think about the tradeoff between RLHF-style preference learning and RLVR (rule-based verifiable rewards) for improving Claude Code's coding quality? What signals would you use to decide which approach to invest in?**\n\n**Why they may ask:** This bridges your RL engineering work to Anthropic's model training decisions. A PM for this role needs to have an informed opinion on how training choices affect the product \u2014 not just the features that sit on top.\n\n**How to prepare:** Prepare a framework: RLVR works well when ground truth is verifiable (test pass/fail, compilation success, linting, type checking \u2014 all natural signals in a coding context); RLHF is needed for subjective quality (code readability, architectural decisions, documentation quality). Reference your Reward Lab work (RLVR, learned, and hybrid reward functions) and connect it to what signals Claude Code could realistically instrument as verifiable rewards. The interesting edge cases \u2014 \"does the code pass tests but is still bad?\" \u2014 are worth raising proactively.\n\n---\n\n### C. Gap and Transition\n\n**Question 9.** **You've been running two AI startups simultaneously since September 2024 \u2014 StreamIO and Fintellect. Neither appears to have significant traction yet. Why join Anthropic as a PM now rather than continuing to build, and what does that say about your risk tolerance and commitment?**\n\n**Why they may ask:** Two concurrent founder roles with limited visible traction, transitioning to a senior IC PM role at a large company, is a narrative that needs a direct answer. Anthropic values intellectual honesty over narrative polish \u2014 a technically sophisticated interviewer will probe the details if you over-spin.\n\n**How to prepare:** Be honest and direct. Frame the decision as a deliberate choice to have maximum impact at the frontier of AI development, not a retreat. Emphasize what you built and learned \u2014 full-stack AI product building, agentic orchestration, eval engineering, MCP integration \u2014 and why Claude Code specifically is where you want to apply it. The strongest version of this answer acknowledges the traction question head-on: \"I built real systems and learned real things; I'm making a deliberate choice to apply that at the scale where it matters most.\" Avoid making the startups sound more successful than they are.\n\n---\n\n**Question 10.** **Your most recent Staff PM role was at Intuit, ending in September 2024. The past 18 months have been founder/builder mode. How do you think about re-entering a large organization's PM structure \u2014 dealing with roadmap dependencies, XFN alignment, and decisions you don't control?**\n\n**Why they may ask:** Anthropic is fast-moving but structured, with research, safety, and engineering teams that all have meaningful input on Claude Code's direction. A founder who's been operating autonomously may struggle with the collaborative constraints of a structured org. This is a legitimate concern, not a gotcha.\n\n**How to prepare:** Acknowledge the shift directly \u2014 don't pretend it's not a real transition. Then draw on your Intuit experience to show you've operated effectively in complex orgs before: working with the CTO on language strategy, XFN alignment across 20+ mobile apps, navigating the ICE platform's multi-team dependencies. Frame your founder experience as making you a *better* collaborator: \"I understand what it takes to ship, so I know how to be a useful XFN partner rather than a blocker.\" The key is showing you've done both modes and can choose the right one.\n\n---\n\n**Question 11.** **You've been an adjunct professor at De Anza since 2018 \u2014 teaching Java, cloud computing, ethical hacking, and data analytics. That's a significant ongoing time commitment. How does that fit with the expectations of a Staff/Senior PM role at Anthropic?**\n\n**Why they may ask:** Anthropic's hybrid policy (25%+ in office) plus the intensity of a Claude Code PM role may conflict with ongoing teaching obligations. This is a practical question, not a philosophical one.\n\n**How to prepare:** Be clear and specific about the actual time commitment \u2014 typically 3\u20136 hours per week per course, concentrated on specific days. Frame teaching as a developer empathy asset that's directly relevant to Claude Code: you understand how developers learn new tools, which informs documentation strategy, onboarding design, and ecosystem building. Be prepared to say clearly that you'd reduce your teaching load if the role required it \u2014 don't leave this ambiguous.\n\n---\n\n### D. Behavioral and Situational\n\n**Question 12.** **Tell me about a time you had to kill a feature or product direction you personally believed in because the data or customer feedback said otherwise. What did you learn?**\n\n**Why they may ask:** Anthropic values empirical thinking and intellectual honesty above conviction. Your founder background means you've been the final decision-maker; this question probes whether you can subordinate your own hypothesis to evidence \u2014 a critical PM skill in a research-driven org.\n\n**How to prepare:** Prepare a specific Intuit or Splunk example where you had a strong hypothesis that was invalidated by data. The ICE platform work \u2014 where you used SQL and BigQuery to prioritize developer pain points across 20 mobile apps \u2014 gives you good material. Be specific about what the data showed, how it contradicted your prior, and how you changed course. The learning is as important as the story: what does this tell you about how you make decisions now?\n\n---\n\n**Question 13.** **Describe a situation where you had to translate a complex AI/ML research advance into a concrete product feature for a non-technical audience \u2014 executives, customers, or partners. How did you do it and what was the outcome?**\n\n**Why they may ask:** The JD explicitly calls out \"translate cutting-edge AI advances into practical developer features\" and \"presenting product strategy to executives.\" This is a core job requirement, not a nice-to-have.\n\n**How to prepare:** The Intuit CTO language assessment story is your strongest option here: you analyzed nine languages across usage data and developer feedback and presented strategic investment recommendations to the CTO. Walk through the arc \u2014 what the technical complexity was, how you framed it for a non-technical executive decision, and what the outcome was. Alternatively, the ICE Presence story ($480K/month invoicing impact) shows you can connect a technical capability to a business outcome that executives care about. DeveloperWeek 2022 speaking experience is also relevant context.\n\n---\n\n**Question 14.** **Tell me about the most technically complex product you've shipped. Walk me through a specific technical decision you made that required you to go deep into the engineering \u2014 not just review a design doc, but actually understand the implementation.**\n\n**Why they may ask:** The JD requires \"at least 1 year as a professional engineer\" and \"deep technical background.\" Anthropic will probe whether your technical depth is real \u2014 whether you can go to the implementation level, not just the architecture diagram level.\n\n**How to prepare:** The rSocket migration at Intuit is your best story here: scaling from 6K to 50K TPS, supporting ~1.5M concurrent connections, sub-25ms TP99. Walk through the actual technical decision at the protocol level: why rSocket over WebSocket or gRPC, what the reactive streams model gave you that alternatives didn't, how you validated the latency target under load. Alternatively, the MSaaS drift detection Java JAR library \u2014 where you wrote the code yourself \u2014 demonstrates hands-on engineering depth. Pick the story where you can go deepest on the implementation specifics.\n\n---\n\n**Question 15.** **Claude Code is used by world-class engineers who will immediately notice quality regressions or missing features. Describe a time you managed a highly technical, opinionated user community and had to make a prioritization call that disappointed some of them.**\n\n**Why they may ask:** Claude Code's user base includes elite developers who are vocal and technically sophisticated. Managing this community requires both product conviction and the ability to maintain trust while saying no.\n\n**How to prepare:** The Splunk Scheduler Service story is strong here \u2014 you delivered a complex microservice in four months while balancing internal partner, third-party developer, and Fortune 500 customer requirements using your RICE framework. Show that you can hold a prioritization decision under pressure from opinionated technical users, explain your reasoning transparently, and maintain trust even when the answer is \"not this quarter.\" The ICE Self-Service platform story (reducing onboarding from weeks to minutes) also shows you understand what opinionated developers actually need versus what they say they want.\n\n---\n\n### E. Role-Specific Scenarios\n\n**Question 16.** **It's your first 90 days as PM for Claude Code. You've done customer interviews, reviewed telemetry, and talked to the engineering team. Walk me through how you'd build your first roadmap \u2014 what framework would you use, what data would you prioritize, and what's the first thing you'd ship?**\n\n**Why they may ask:** This is the canonical \"show me you can do the job\" question. Anthropic wants to see your product process, your instinct for developer tooling, and whether you have a specific, informed point of view on Claude Code's current gaps \u2014 not a generic PM framework.\n\n**How to prepare:** Use Claude Code extensively before every interview round \u2014 build something real with it and identify three specific gaps or opportunities (e.g., CLAUDE.md ecosystem discoverability, multi-repo context management for large monorepos, eval/testing integration for agentic tasks). Frame your roadmap process in three layers: qualitative interviews with power users to surface pain points, quantitative analysis of task completion rates and abandonment signals, and research alignment on the model capability roadmap (what's coming in the next 6 months that changes what's possible?). Have a specific first ship with a clear success metric ready \u2014 not a category, but a feature.\n\n---\n\n**Question 17.** **Claude Code is competing with GitHub Copilot, Cursor, Windsurf, and Devin. How would you define Claude Code's differentiated positioning, and what's one feature you'd ship in the next quarter that would widen the moat?**\n\n**Why they may ask:** The JD says Claude Code should remain \"ahead of model capabilities and seen as the best way to experience the most intelligent Claude models.\" You need competitive awareness and a specific product instinct for differentiation \u2014 not generic positioning language.\n\n**How to prepare:** Research the current competitive landscape deeply before the interview. Claude Code's structural moat is model intelligence plus agentic depth plus safety \u2014 but that's the category answer. The interesting answer is: what does Claude Code do that *only* Anthropic can do? Prepare a specific feature idea that leverages Anthropic's unique assets \u2014 extended thinking for complex multi-file refactoring, MCP ecosystem depth as a platform lock-in mechanism, safety-aware code review that flags security vulnerabilities with the same rigor Anthropic applies to Claude's outputs. Have a crisp, defensible positioning statement ready and be prepared to argue for it.\n\n---\n\n### F. Motivation and Fit\n\n**Question 18.** **Anthropic's mission is AI safety \u2014 building reliable, interpretable, and steerable AI systems. How does that mission connect to your personal motivations, and how would it shape how you'd make product decisions for Claude Code specifically?**\n\n**Why they may ask:** Anthropic screens hard for genuine alignment with AI safety values. Your background is heavily capability-focused \u2014 RL workbench, multi-agent orchestration, LLM pipelines. You need to demonstrate that safety is a genuine value, not interview polish. Anthropic interviewers will detect performative safety talk immediately.\n\n**How to prepare:** Prepare a specific answer that connects safety to concrete Claude Code product decisions \u2014 not a values statement. For example: how would you think about agentic task boundaries and permission scoping when Claude Code is operating autonomously on a production codebase? What eval gates would you require before shipping a new agentic capability? Reference your aeval adversarial safety testing work (refusal detection, safety gates in CI/CD) as evidence that safety-conscious engineering is already part of how you build. Be genuine \u2014 if there's a specific safety concern about agentic coding tools that you've thought about, name it.\n\n---\n\n**Question 19.** **You've built two AI startups, published at NeurIPS, taught college courses, and held Staff PM roles at Intuit and Splunk. Why Claude Code specifically \u2014 not Claude's API, not enterprise, not a different AI company?**\n\n**Why they may ask:** Anthropic wants PMs who are genuinely obsessed with the specific product. Your breadth is a strength, but it can read as \"excited about AI generally.\" The answer needs to be specific to Claude Code, not to Anthropic or AI broadly.\n\n**How to prepare:** Prepare a specific answer grounded in a real Claude Code experience. Claude Code is the product where model intelligence, developer tooling, and agentic workflows converge \u2014 which is exactly the intersection of your RL workbench, OpenClaw, and Intuit SDK work. But don't just say that. Have a concrete anecdote: a specific moment when you were using Claude Code and thought \"this is where I need to be\" \u2014 or a specific gap you noticed that you have a clear idea for how to fix. The more specific and personal the answer, the more credible it is.\n\n---\n\n### G. Frontier Thinking\n\n**Question 20.** **You've built your own multi-agent framework and benchmarked RL training frameworks. If you were advising the Claude Code team on what the \"ultimate form factor of agentic software development\" looks like \u2014 the JD's exact phrase \u2014 what would you say, and what's the biggest open research question that needs to be solved to get there?**\n\n**Why they may ask:** Anthropic's JD explicitly says \"the ultimate form factor of agentic software development remains unwritten.\" This is an invitation to think at the frontier. Your combination of RL workbench engineering, multi-agent orchestration, and eval platform work makes you unusually qualified to have a specific, grounded opinion \u2014 not just a vision statement.\n\n**How to prepare:** Prepare a 2\u20133 minute answer with a specific thesis. For example: \"The ultimate form factor is a persistent, context-aware agent that owns a codebase the way a senior engineer does \u2014 with long-term memory across sessions, judgment about when to act versus when to ask, and verifiable correctness guarantees before committing changes.\" Then name one open research question that's genuinely hard: long-horizon task decomposition with reliable error recovery, or formal verification of agentic code changes at scale, or the context window economics of maintaining codebase-level understanding. Ground it in what you learned building OpenClaw and aeval \u2014 not just what you've read.\n\n---\n\n**Question 21.** **Anthropic recently acquired Stainless and is clearly investing in the developer platform layer. If you were scoping the Claude Code PM role to include SDK ecosystem strategy \u2014 not just the CLI \u2014 what would the 3-year vision look like, and what's the first platform primitive you'd build?**\n\n**Why they may ask:** The Stainless acquisition is a major signal about Anthropic's developer platform ambitions. Your Intuit experience (SDK Starter Kits, DevPortal, GitOps, ICE Playground) is the most directly relevant background in your r\u00e9sum\u00e9 for this strategic question.\n\n**How to prepare:** Prepare a staged vision: Year 1 \u2014 MCP server registry and Claude Code plugin ecosystem with standardized CLAUDE.md schema and community sharing; Year 2 \u2014 Stainless-powered auto-generated SDKs for any API integrated via Claude Code, so developers can go from \"Claude Code found this API\" to \"Claude Code generated a typed client\" in one step; Year 3 \u2014 Claude Code as the default development environment for building on Anthropic's platform, with first-class SDK support for every Anthropic product. First primitive: a standardized, versioned CLAUDE.md schema with community discovery and usage signals. Draw explicit parallels to your Intuit DevPortal work \u2014 you've built this category of product before.\n\n---\n\n### H. Product Prioritization and Metrics\n\n**Question 22.** **You have three competing Claude Code initiatives: (1) improving multi-repo context management for large codebases, (2) building a community ecosystem for sharing CLAUDE.md configurations and slash commands, and (3) deeper IDE integration beyond VS Code. You can only ship one in Q3. How do you decide?**\n\n**Why they may ask:** The JD calls out roadmap definition and ecosystem building as core responsibilities. Anthropic wants to see a rigorous, data-driven prioritization process \u2014 not gut instinct dressed up as a framework.\n\n**How to prepare:** Apply a structured framework and walk through it explicitly. Define your decision criteria first: reach (how many active users does this affect?), impact on power users versus new users (which matters more for Claude Code's current stage?), strategic moat (which widens the gap from competitors?), engineering feasibility (what's the actual cost?). Then walk through each option against those criteria. Be prepared to defend your choice with specific reasoning and to acknowledge what you're giving up. Reference your Splunk RICE framework experience and explain how you'd adapt it for Claude Code's developer-first context \u2014 where power user retention may matter more than raw reach.\n\n---\n\n**Question 23.** **What's the north star metric for Claude Code, and what are the 3 leading indicators you'd track to know if you're on track to hit it? What counter-metrics would you watch to make sure you're not gaming the north star?**\n\n**Why they may ask:** This probes whether you can design a metrics system that actually measures what matters \u2014 not just what's easy to count. Your telemetry work at Intuit (SQL, BigQuery, 20 mobile apps) and your aeval statistical rigor make this directly testable.\n\n**How to prepare:** Propose a specific north star and defend it: \"Tasks completed autonomously per active developer per week\" captures both adoption and agentic quality \u2014 it goes to zero if developers don't use Claude Code, and it goes to zero if Claude Code can't complete tasks without constant intervention. Leading indicators: daily active agentic sessions (are developers starting tasks?), task completion rate without human intervention (is the agent actually working?), time-to-first-successful-task for new users (is onboarding working?). Counter-metrics: task abandonment rate, error recovery loop count, user override frequency. Show you understand the difference between vanity metrics and leading indicators \u2014 and that you've thought about how a metric can be gamed.\n\n---\n\n## 3. Preparation Priorities\n\n1. **Use Claude Code deeply and daily before every interview round.** Build something real with it. Have specific, opinionated feedback on what's broken, what's great, and what you'd change. Generic enthusiasm will not pass Anthropic's bar \u2014 you need a product POV grounded in hands-on experience. Identify three specific gaps and have a prioritized first-ship ready with success metrics.\n\n2. **Prepare your technical credibility stories with empirical specificity.** The RL workbench, OpenClaw, and aeval are your strongest differentiators \u2014 but only if you can go beyond architecture descriptions to concrete findings. What surprised you? What failed? What did you measure? Anthropic's research team will probe for depth, not breadth.\n\n3. **Nail the founder transition narrative.** Two concurrent startups with limited visible traction transitioning to a Staff PM role is the most obvious flag in your profile. Prepare a crisp, honest, forward-looking answer: what you built, what you learned, and why Claude Code is a deliberate choice \u2014 not a fallback. Intellectual honesty will land better than narrative polish with this audience.\n\n4. **Develop a specific Claude Code roadmap thesis.** Research the current product deeply: CLAUDE.md, MCP servers, slash commands, VS Code extension, CLI behavior. Identify three specific gaps. Have a prioritized first-ship ready with a clear success metric. The JD says \"define the roadmap\" \u2014 show up with a draft, not a process description.\n\n5. **Connect your work to Anthropic's safety mission with specificity.** Your background is capability-heavy. Prepare concrete examples of how you'd apply safety thinking to Claude Code product decisions: agentic task boundaries, permission scoping, user control, eval gates before shipping new agentic capabilities. Reference your aeval adversarial safety testing as evidence this isn't performative.\n\n---\n\n## 4. Watch-Outs\n\n**Founder traction gap.** StreamIO and Fintellect have been running since September 2024 with no visible traction metrics in the r\u00e9sum\u00e9. If asked about outcomes, don't deflect or over-spin. Anthropic values intellectual honesty, and a technically sophisticated interviewer who probes the details will see through narrative inflation immediately.\n\n> Instead: \"We're early-stage \u2014 here's what I built, here's what I learned about X, and here's why I'm making this deliberate choice now rather than continuing to build.\"\n\n**Breadth versus depth perception.** Your r\u00e9sum\u00e9 spans RL workbenches, multi-agent frameworks, eval platforms, protein structure prediction, real estate APIs, financial education apps, drone licenses, saxophone, and triathlon. This can read as scattered rather than focused.\n\n> Proactively frame the narrative: everything connects to \"building AI systems that developers can trust and use effectively.\" Have a 30-second version of this framing ready for any question that touches on your background.\n\n**PM versus builder identity confusion.** You've done significant hands-on engineering \u2014 FFmpeg pipelines, Java JAR libraries, React/TypeScript apps. In a PM interview, this is a strength \u2014 but only if you can clearly articulate the *product decisions* you made, not just the code you wrote.\n\n> For every technical story, be ready to answer: \"What was the product decision here, and how did you make it?\" Don't let technical depth crowd out the PM narrative.\n\n**Safety performativity.** Anthropic screens hard for genuine AI safety alignment. If your safety answers sound like \"I care about safety because Anthropic cares about safety,\" it will land badly.\n\n> Prepare specific, grounded safety thinking: how would you handle a Claude Code feature that could be used for malicious code generation? What eval gates would you require before shipping an agentic capability that can modify production files? Ground answers in your aeval adversarial testing work and your actual product decisions \u2014 not values statements.\n\n---\n\n## 5. Questions to Nail\n\nThese are the five highest-stakes questions to rehearse. A strong answer to each one is what separates a compelling candidate from a forgettable one.\n\n**1. The RL workbench empirical findings question (Question 2)**\nThe one thing a strong answer must land: *specific, surprising empirical findings with numbers* \u2014 not a description of the architecture. If you can't name what surprised you, the workbench reads as a r\u00e9sum\u00e9 project rather than real research.\n\n**2. The OpenClaw vs. Claude Code comparison (Question 3)**\nThe one thing a strong answer must land: *a specific, grounded critique of Claude Code's current agentic approach* based on what you learned building OpenClaw. This proves you've used the product deeply and have genuine product insight.\n\n**3. The founder transition question (Question 9)**\nThe one thing a strong answer must land: *intellectual honesty plus a clear, forward-looking rationale* for why Claude Code specifically is where you want to apply what you've built. Avoid defensiveness and avoid over-spinning traction.\n\n**4. The 90-day roadmap question (Question 16)**\nThe one thing a strong answer must land: *three specific, named Claude Code gaps with a prioritized first-ship and a success metric.* A process description without a specific product opinion will not pass the bar.\n\n**5. The ultimate form factor question (Question 20)**\nThe one thing a strong answer must land: *a specific thesis about what agentic software development looks like at its best, plus one genuinely hard open research question.* This is your chance to demonstrate frontier thinking \u2014 don't give a vision statement, give a position.\n\n---\n\n## 6. Suggested Narrative\n\nThe through-line Felix should reinforce across all answers is this: **I have been building the components of what Claude Code is trying to become \u2014 from the eval infrastructure to the multi-agent orchestration to the developer platform tooling \u2014 and I'm joining Anthropic to do it at the scale and with the model intelligence that makes it matter.**\n\nThis narrative works because it's true and it's specific. The RL workbench is not a side project \u2014 it's the kind of empirical research Anthropic's team does. OpenClaw is not a toy \u2014 it's a production multi-agent orchestration system built on top of Claude's MCP SDK. aeval is not a dashboard \u2014 it's a principled evaluation platform with statistical rigor that mirrors how Anthropic thinks about model quality. And the Intuit DevPortal and SDK Starter Kit work is directly analogous to what Anthropic needs to build on top of the Stainless acquisition.\n\nThe risk to this narrative is breadth \u2014 it can sound like \"I've done everything.\" Anchor it with specificity: \"The specific thing I've built that maps most directly to Claude Code's trajectory is X, and here's what I learned that I'd apply immediately.\"\n\n---\n\n## 7. Final Reminder\n\nWalk into every Anthropic interview prepared to demonstrate:\n\n- **Genuine hands-on Claude Code use** \u2014 specific opinions, specific gaps, specific ideas, not general enthusiasm\n- **Empirical depth** \u2014 findings, numbers, and honest accounts of what failed, not just architecture descriptions\n- **PM decision-making** \u2014 for every technical story, the product decision and how you made it, not just the implementation\n- **Safety thinking that's grounded** \u2014 specific product decisions, not values statements\n- **Intellectual honesty** \u2014 about the startups, about the transition, about what you don't know\n- **Specificity about Claude Code** \u2014 why this product, not just why Anthropic or why AI\n- **The builder-PM identity** \u2014 you are a technical builder who makes product decisions, and every answer should reinforce both halves of that",
  "source_artifact_id": "art_gL-gR72mefI",
  "source_kind": "interviewer_questions",
  "doc_type": "questions_they_ask"
}