Anthropic Engineering Director: 5 Process Overhauls for AI-Native Teams
AI Summary · From a Serial Entrepreneur’s Perspective (The content below is distilled by AI; the views belong to the original author. You can read this without visiting the source.)
Fiona Fung, Director of Anthropic’s Claude Code team, shares how her organization restructured after shifting from a “coding bottleneck” to a “verification bottleneck.” The core changes include moving planning from six-month roadmaps to JIT prototype iterations; having humans review only domain-specific security while AI handles routine code reviews; and hiring exclusively for creative builders and systems experts. Metrics from hands-on practice show new hires shipping real code within their first week.
- Replacing pre-planning with JIT planning…
- Layered code reviews: style and tests go to AI; humans judge only security boundaries and product decisions
- Focusing hiring on two profiles: product-savvy creative builders + deep systems experts…
- Tracking three metrics: Onboarding time…
- Team principles: everyone dogfoods AI tools; managers start as individual contributors…
From a “Code Bottleneck” to a “Verification Bottleneck”: Engineering Restructuring at Anthropic’s Claude Code Team
The underlying assumptions of engineering organizations are failing. For decades, whether using waterfall or agile, the premise has been the same: writing code is expensive and scarce. But at Anthropic’s Claude Code team, Fiona Fung found the bottleneck had already shifted—code generation no longer blocks progress. Verification, review, and security assessment have become the new chokepoints.
Key takeaway: Don’t try to fit old processes into new tools; rewrite the processes themselves. Below are the five key dimensions they restructured and the落地 metrics they track.
1. Planning: From Six-Month Roadmaps to JIT Prototype Iteration
Code used to be expensive, so teams invested heavily in upfront planning. Now change happens too fast for a six-month roadmap to survive three months. The team switched to JIT (Just-In-Time) planning:
- What they do: Skip long design docs. Ship a prototype first, let internal users try it, then iterate based on feedback.
- Ceremony shift: Fewer product review meetings. Discussions inside PRs now serve as the planning artifact.
2. Code Reviews: Layered Handling, Humans Judge Only “Domain Safety”
Claude now handles all style checks, linting, bug catching, and test generation. Human reviewers’ authority shrinks to three areas:
- Legal and security boundaries: Code touching data privacy or compliance.
- Product judgment and taste: UX details only a human can decide.
- Trust boundaries: Core architecture decisions.
Note: As models improve, this boundary shifts right dynamically. What requires human review today may be fully automated next month.
3. Context Acquisition: Ask Claude, Not the Author
When AI helps generate code, “who wrote it” no longer matters. The new knowledge path is:
- Ask Claude first: It combines commit history, docs, and the codebase to give richer context.
- Then ask “Can this be automated?”: For repetitive questions (like daily customer feedback summaries), write an automated job instead of pestering a person.
4. Team Composition: Hire Only Two Types of People
Role boundaries are blurring (PMs write code; engineers do design). When hiring, focus on two traits—not raw output speed, which is AI’s strength:
- Creative Builders: Deeply curious, product-minded, eager to solve real problems with tech.
- Deep Systems Experts: Strong accumulated knowledge in specific domains (e.g., web performance, infrastructure) to handle complex system issues AI can’t tackle.
5. How to Land It and Measure
Three hard rules:
- Everyone dogfoods: All team members (including managers) must use Claude Code/Cowork. Managers start as ICs (individual contributors), shipping code to understand the engineering reality.
- Flat structure: One team mission; people freely flow to pods with work.
- Kill empty processes: Any process that no longer creates value can be canceled on the spot by any team member (e.g., axing useless weekly syncs).
Three tracking metrics (start now):
- Onboarding time: Can a new hire ship real code in week one? (The Claude Code team has reached this standard.)
- PR cycle time: If code volume surges but cycle time doesn’t drop, CI/CD or review is dragging you down.
- Share of commits assisted by Claude: Target close to 100%.
Watch out: Throughput ≠ success. Make sure AI-generated code actually solves business problems, not just increases line counts.
6. Action Guide: Start with the Noisiest Workflow
If you don’t know where to begin, ask yourself: What’s the most expensive, headache-inducing, least-desired part of your workflow?
Then follow up: Is it still serving its purpose? If so, can it be automated or cut?
For example, if during weekly syncs everyone stares at their laptops and only looks up to report status, cancel the meeting. Find that pain point and restructure it with AI—don’t patch it.
Original post · Steve Sun: Read the original →
Related tool recommendation (promoted): GLM Coding Plan — AI Coding Powered by G…