Why Claude Code Hires Only Dreamers and Deep System Experts

Core Insight

The team’s premise is that traditional engineering processes were built on the assumption that "writing code is expensive"; now that AI makes code cheap, those processes become obsolete, causing a "shift" of bottlenecks to verification, code review, and security.

The Shift Across Eras

2000s: Software shipped on CD‑ROMs with hard deadlines (e.g., Visual Studio 2005 at Microsoft).

Internet era: Online distribution enabled continuous updates and agile development.

Agentic coding era: Code is no longer the bottleneck; old processes need a new re‑arrangement.

Five Process Changes (Ordered by Importance)

Planning – moving from a six‑month roadmap to Just‑In‑Time (JIT) planning.

Code Review – redefining the division of labor between Claude and humans.

Onboarding – ensuring new hires ship real code within the first week.

Recruiting & Team Composition – focusing on two talent archetypes.

Organizational Shape – flattening hierarchy and aligning on a single mission.

JIT Planning

Instead of a static design doc, the team prototypes directly, skips formal product reviews, and conducts discussions in pull‑requests.

A story illustrates the change: when a teammate wanted to refactor, they asked Claude to generate three different PRs, instantly seeing the trade‑offs and impact on callers, turning a traditional debate into a data‑driven comparison.

Code Review: Trust but Verify

Claude automatically handles style/lint, PR feedback, common bugs, and test generation. Humans retain responsibility for legal review, security‑sensitive code, and product sense.

"Code is cheap, but the culture of alignment becomes more valuable."

Hiring Profile

The team looks for only two types of people:

Creative builders with product sense: Dreamers who instantly envision a product from a problem and iterate relentlessly.

Deep systems experts: Specialists in distributed systems needed to run Claude Code on the web.

They no longer evaluate candidates by raw commit volume; instead they assess judgment, product intuition, and low‑level understanding.

Organizational Shape

Flat teams where managers must "dog‑food" (write code themselves) to retain credibility. A unified mission avoids fragmented pod goals and enables rapid pivots.

Metrics for Success

Onboarding ramp‑up time: Should decrease; new hires ship real code in the first week.

PR cycle time: Should decrease; an increase signals CI bottlenecks.

Claude‑assisted commit ratio: Should increase; the team observed four months with no non‑Claude commits.

Beware of Mis‑leading KPIs

"Don't confuse throughput with success."

AI‑generated code volume is a metric, not a goal. Focusing on it can create perverse incentives to make AI write unnecessary code.

Open Questions

Should iOS and Android teams still be separate when engineers can cross‑platform effortlessly?

To what extent should review automation be pushed before important checks are missed?

How to keep every role feeling valuable when responsibilities blur?

Final Reflections

The hardest part isn’t adopting AI but killing legacy processes. Explicit permission to discard outdated workflows is essential for teams transitioning to AI‑native development.