Lenny's PodcastWhat happens after coding is solved? | Fiona Fung (Claude Code & Cowork)
CHAPTERS
Fiona Fung’s journey: from IBM terminals to Visual Studio dogfooding
Lenny introduces Fiona Fung and frames the conversation around the dramatic productivity jump from AI-assisted coding. Fiona recounts early career shifts—IBM’s low-level systems work and Microsoft’s IDE revolution—and how dogfooding shaped her product instincts.
- •AI has driven an enormous jump in engineering throughput (Anthropic’s 8x code shipped metric)
- •Early-career tooling shifts (terminal/Vim → IDEs) were major step changes in productivity
- •Dogfooding at Visual Studio created fast feedback loops even before social media
- •Shipping constraints evolved from CD releases to online delivery, changing planning and iteration cycles
When coding stops being the bottleneck: verification becomes the new constraint
Fiona explains why the core bottleneck has moved from writing code to verifying correctness and impact. The team’s challenge is less about generating output and more about ensuring quality, reliability, and user experience across fast-moving surfaces.
- •Coding output is abundant; verifying quality and impact is the limiting factor
- •More disciplines (PMs, designers) can now ship code, raising the bar for shared verification systems
- •Teams must rethink quality practices for much higher change velocity
- •Trust-but-verify becomes the guiding operating principle
What an “AI‑pilled” team looks like in 2026: roles blur into builders
Fiona describes the emerging team model where nearly everyone becomes a “builder,” and traditional role boundaries fade. Management practices shift toward enabling autonomy, scaling feedback, and maintaining clarity on outcomes.
- •Everyone trends toward being a builder; roles blur across engineering/PM/design
- •Leaders focus on enabling throughput plus verification, not just task assignment
- •Outcome orientation matters more than job titles or traditional ownership boundaries
- •High velocity requires new team rituals and tooling to stay aligned
Claude as a manager’s operating system: visibility, retros, and coaching via a shared agent
Fiona details how she uses Claude Code as a persistent session with access to repos, Slack, and metrics to review team output and drive better conversations. Monthly look-backs become richer: what shipped, how it performed, what broke, and what to improve.
- •A persistent Claude Code session provides cross-repo visibility and context
- •Monthly reviews focus on impact, feedback, incidents, and investment themes
- •Claude helps synthesize patterns across incidents to identify quality hotspots
- •Management shifts from “tracking work” to “making better decisions faster”
From manual feedback sweeps to automated routines: the rise of async agent workflows
Fiona explains how “routines” (cron-like automation) changed her daily ritual of scanning feedback channels. Instead of summarizing issues manually, routines can summarize themes and even propose PRs, pushing work toward an asynchronous, agent-driven operating model.
- •Routines run on schedules to monitor channels and generate actionable summaries
- •Agents can pre-draft fixes and PRs, reducing time-to-response on feedback
- •Async workflows raise abstraction: agents generate prompts, not just answers
- •The future trend is toward fleets of agents operating in the background
Code review evolves: frameworks, specs-in-repo, and AI validation against “what good looks like”
With human review becoming a bottleneck, Fiona emphasizes encoding standards and expectations into repos so AI can review against them. This extends test-driven thinking and makes quality checks more scalable as output explodes.
- •Claude-assisted code review didn’t exist a year earlier; now it removes a major bottleneck
- •Best results come from explicit frameworks/specs checked into the repo
- •AI excels at validating work against written expectations and standards
- •TDD principles get easier when tests and scaffolding are generated automatically
Hiring for the AI era: creative product builders + deep systems experts
Fiona shares the two profiles Anthropic prioritizes: end-to-end builders with product sense and deep subject-matter experts for critical systems. AI increases what generalists can do, but verification and hard infrastructure still demand real expertise.
- •Creative builders excel at ideation, iteration, polish, and owning outcomes
- •Deep systems experts remain essential for distributed systems and hard verification areas
- •Teams must identify where expert human judgment is non-negotiable
- •AI “lifts the ceiling,” enabling engineers to work across domains (e.g., mobile)
Ambition, growth mindset, and fear: who thrives vs. who resists
The conversation shifts to mindset: ambitious thinking becomes the differentiator when “everything is possible in theory.” Fiona highlights growth mindset and leaning into fear as key traits for adapting, and shares personal stories about taking control amid uncertainty.
- •AI shifts constraints from feasibility to ambition and prioritization
- •Growth mindset: what made you successful before may stop working
- •Resistance often hides fear; focusing on controllable actions reduces anxiety
- •Doing “something scary” periodically is framed as a path to growth
Bridging the AI adoption divide: helping small businesses and spreading practical use cases
Fiona explains her passion for small businesses and how Cowork can remove painful admin work like invoicing and expenses. She argues adoption spreads best through concrete, relatable examples and hands-on help, especially for those hesitant about AI.
- •Small businesses can gain outsized leverage from automating paperwork and analysis
- •Real-world onboarding reveals product bugs and unexpected use cases (e.g., menu discovery, pricing analysis)
- •Sharing personal, practical examples is an effective adoption catalyst
- •Equitable access matters to prevent widening capability gaps
How Anthropic spots latent demand: watching “people jumping through hoops”
Fiona describes the product pattern of noticing unexpected usage and turning it into a first-class experience. Latent demand emerges when users hack around limitations; teams should form hypotheses and smooth the workflow rather than fight the behavior.
- •Latent demand signals appear when users use products in unintended but valuable ways
- •Teams should ask: what are users trying to accomplish, and where are the hoops?
- •Rapid iteration and tight feedback loops are central advantages
- •Claude for small business is cited as bundling needs discovered through user behavior
Agency with accountability—and the shift from token maxing to outcome/ROI measurement
Fiona emphasizes a team culture of high agency paired with high accountability, grounded in hypotheses and measurable outcomes. She critiques shallow productivity metrics and argues for aligning outputs to real outcomes, using metrics as tools—not targets.
- •High agency works only when paired with clear accountability and hypotheses
- •Measuring productivity (LOC, PRs, tool usage) is often misleading or gameable
- •Focus on outcomes: avoid “motion vs. progress” traps
- •Leaders should do listening tours with senior engineers to learn what’s actually working
Quality at scale: proactive monitoring and the “bad vs. sad” framework
To manage quality amid high velocity, the team invests in proactive detection and shared language. “Bad vs. sad” helps teams classify irrecoverable failures vs. recoverable pain points, enabling consistent prioritization across disparate surfaces.
- •Proactive detection and monitoring matter more as throughput rises
- •“Bad” = irrecoverable failures (e.g., crashes, lost work); “sad” = recoverable friction (e.g., flicker)
- •Teams define bad/sad per surface, preserving agency while standardizing language
- •Stacked “sads” can become “bad,” so trend monitoring is crucial
Why managers start as ICs: dogfooding, credibility, and preventing skill atrophy
Fiona explains Anthropic’s preference for managers to begin as individual contributors and continue hands-on work. Staying in the flow improves product intuition, builds rapport, and helps leaders understand rapidly changing tools and codebases.
- •Starting as an IC lets new managers learn the product/code without “managery” reflexes
- •Hands-on dogfooding keeps leaders grounded in real user experience beyond dashboards
- •Claude makes it easier for long-time managers to safely ship code again
- •Skill atrophy concerns remain: engineers should still understand architecture and dependencies
What’s lost (and what’s next): loneliness, context switching, changing PM/data science, and JIT planning
They discuss tradeoffs: less classic “flow,” more loneliness, and heavier context switching with many async agents. Fiona highlights evolving PM and data science roles, open questions about automated review and org structure, and a move from roadmaps to monthly just-in-time planning.
- •Agents reduce solo grind but can increase loneliness; team rituals like pair-programming lunches help
- •Context switching grows with many parallel agents; solutions are still unclear
- •PMs and other roles increasingly ship code; data science shifts toward review/verification of AI-driven analysis
- •Planning moves from 6-month roadmaps to monthly JIT priorities with frequent re-checks
- •Open questions: mobile org structure, how far to automate reviews, and how to ensure productivity amid role blurring
