Building Is Cheap, Arguing Is Expensive

Sources#

• Running an AI-native engineering org

Summary#

Fiona Fung's rule for technical debates once agentic coding makes generation nearly free: "in technical debate, code wins — building is cheap, argument is expensive." When you can generate three real implementations in the time it takes to book a whiteboard session, you stop arguing about approaches and start comparing them. The corollary norms on Claude Code: prototype-and-dogfood instead of writing design docs, and reduce in-depth planning docs in favor of discussion in PRs and prototypes.

The generate-three-PRs move#

Fung's war story: onboarding to Claude Code, she and Boris Cherny debated a refactoring approach. Old reflex: "let's go to a meeting room and whiteboard it." Instead: "thanks to Claude, I generated three different versions of the PRs" — which gave them a better technical debate, because she could compare not just how the code was implemented but the impact to the callees. The artifact (real, runnable PRs) carries more information than the argument (whiteboard abstractions), and producing it is now cheap.

Prototype over design doc#

The same economics rewrite prototyping. Previously prototyping split teams into two camps — fast feedback vs. cutting corners and getting stuck shipping throwaway code that wasn't built to scale. With Claude, the second objection weakens: "prototyping is a great way to get started because we can iterate and learn, and then scale the prototype to production a lot faster." So the Claude Code norm is: have an idea → prototype it → dogfood it → then decide. (Cf. the prototype "throwaway to flesh out a design" approach.)

What gets reduced: design docs#

The thing the team reduced: "really in-depth planning and design docs. Most of our discussions actually happen in PRs or prototypes." This is downstream of Verification as the New Bottleneck — when engineering bandwidth is no longer scarce, the heavy upfront planning that existed to protect that bandwidth loses its justification. Refactoring and architecture cleanup, which used to lose the "when will we find time?" budget fight, also stop being a bottleneck.

Tension to hold#

This norm is in productive tension with the wiki's planning-first concepts. Design Concept Grilling (Matt Pocock) and Vertical Slice Tracer Bullets argue you must reach a shared design concept before building. The reconciliation: "building is cheap" doesn't abolish design thinking — it relocates it. You still need taste and a spec (Outsource Your Thinking, Not Your Understanding), but you arrive at them by comparing built artifacts rather than by arguing in the abstract. Cheap building makes the prototype the medium of design, not a replacement for design. (Karpathy strikes the same balance in Vibe Coding vs. Agentic Engineering: the human still owns the spec.)

Connections#

• Fiona Fung — "building is cheap, argument is expensive"
• Boris Cherny — the generate-three-PRs refactoring debate partner
• Verification as the New Bottleneck — the upstream cause: generation cheap → planning-to-protect-bandwidth loses its point
• Design Concept Grilling — the productive tension: design-before-build vs. build-to-decide (reconciled: prototype is the design medium)
• Vertical Slice Tracer Bullets — tracer bullets are the "build a thin real thing to learn" version of the same instinct
• Building Is Cheap, Arguing Is Expensive — see also Disposable Micro-Apps: throwaway built artifacts as the unit of decision
• Disposable Micro-Apps — Thariq Shihipar's throwaway UIs are the same "cheap built artifact beats abstract argument" logic, applied to plan-editing
• Outsource Your Thinking, Not Your Understanding — building is cheap, but the human still supplies the taste/spec that judges the three PRs
• AI Native Product Cadence — prototyping-over-docs is part of the lighter-PRD cadence

Open Questions#

• When does "generate three and compare" become wasteful — at what decision weight is a real argument (or a design doc) still cheaper than three implementations?
• If design discussion lives in PRs/prototypes, where is the rationale recorded for future readers — does the "why we chose this" knowledge survive, or does it share the staleness problem of Code as Source of Truth?

Sources#

• Running an AI-native engineering org