Vibe coding in prod | Code w/ Claude

AAnthropicYouTube channel

Video by Anthropic on YouTube · Summary in English · 31:13 video · Added Apr 29, 2026

TL;DR

Eric from Anthropic discusses 'vibe coding'—fully trusting AI to generate code without reviewing it—and argues it can be used safely in production by forgetting the code but not the product. He emphasizes acting as a product manager for AI, focusing on leaf nodes, and designing verifiable checkpoints. With AI capabilities doubling every seven months, engineers must adapt to avoid becoming bottlenecks.

Key points

Vibe coding means fully trusting AI to write code without reviewing it, as defined by Andrej Karpathy.
The exponential growth of AI capabilities (doubling every seven months) will soon make manual code review impractical.
To vibe code safely, forget the code but not the product; verify at a higher abstraction layer.
Focus AI-generated code on leaf nodes (end features with no dependencies) to contain tech debt.
Act as a product manager for Claude: provide detailed requirements, specs, and constraints upfront.
Use test-driven development with simple end-to-end tests to verify correctness without reading code.
Real-world example: Anthropic merged a 22,000-line change written by Claude using these principles.
Non-technical users should not vibe code production systems; they lack the context to guide AI effectively.
The industry must embrace exponentials: AI will be millions of times better over time, not just twice as good.
Learning accelerates with AI; engineers can take more 'shots on goal' and learn from more iterations.

Detailed summary

Speaker Introduction & Context

Eric, a researcher at Anthropic focused on coding agents, co-authored 'Building Effective Agents' with Barry Zang. He shares a personal anecdote: after breaking his hand, Claude wrote all his code for two months, which inspired his research and product work.

Definition of Vibe Coding

Vibe coding, attributed to Andrej Karpathy, is 'where you fully give into the vibes, embrace exponentials, and forget that the code even exists.' The key is forgetting the code exists. This became significant when non-engineers could build apps without coding knowledge.

Downsides of Vibe Coding

Random bugs, maxed API keys, subscription bypasses, database corruption.
Successful cases are low-stakes: games, side projects where bugs are acceptable.

Why Vibe Coding Matters for Production

The exponential: AI task length doubles every seven months. Currently ~1 hour; next year ~1 day; the year after ~1 week. Engineers cannot review every line at this pace. The industry must find ways to responsibly give into this.

Analogy: Compilers

Early compilers were distrusted; developers read assembly output. This doesn't scale. Eventually, trust is necessary. The question is how to do it responsibly.

Core Thesis: Forget the Code, Not the Product

Eric challenges: 'How will we vibe code in prod and do it safely?' Answer: 'We will forget that the code exists but not that the product exists.' This is analogous to compilers—we know assembly exists but don't think about it.

This is Not a New Problem

Managing implementations you don't understand is age-old. Examples: CTOs use acceptance tests; PMs use the product; CEOs spot-check data. Software engineers must let go of full-stack understanding to become most productive.

Caveat: Tech Debt

Currently, no good way to measure tech debt without reading code. So vibe coding must be targeted.

Strategy: Focus on Leaf Nodes

Leaf nodes (end features with no dependencies) can have tech debt. Core architecture (trunks and branches) must be deeply understood. Models are improving (Claude 4), expanding the trust boundary.

How to Succeed: Be Claude's PM

'Ask not what Claude can do for you, but what you can do for Claude.' Provide guidance like a PM: codebase tour, requirements, specs, constraints. Eric spends 15-20 minutes collecting guidance into a single prompt, then has Claude execute. High success rate.

Who Should Vibe Code?

Not for non-technical people building businesses from scratch; they can't ask the right questions.

Real-World Example: 22,000-Line Change

Anthropic merged a 22,000-line change to production RL codebase written heavily by Claude. How:

Acted as PM: days of human work on requirements.
Focused on leaf nodes: tech debt acceptable there.
Designed stress tests with verifiable inputs/outputs. Outcome: same confidence, fraction of time. Feature that cost two weeks now costs one day.

Closing Thoughts

Be Claude's PM.
Focus on leaf nodes.
Think about verifiability.
Remember the exponential: not vibe coding today is okay, but in a year or two, reading every line will be a bottleneck.

Q&A Highlights

Learning: AI enables faster learning; lazy people won't learn, but curious ones can use Claude to understand code.
Pre-Planning: Provide detail based on what you care about; don't over-constrain.
Security: Know enough to guide Claude; high-profile failures come from non-developers building production systems.
Accessibility: Excited for provably correct frameworks (e.g., backend built for you, UI only) and safe sandboxes like Claude Artifacts.
TDD: Very useful; write simple end-to-end tests. Eric often reads tests first.
Embracing Exponentials: Models will be millions of times better over 20 years, not just twice as good.
Workflow: Uses Claude Code and Cursor; compacts sessions at good stopping points.

Additional Q&A (Part 2)

Tool Usage: Uses Claude Code and Cursor; starts with Claude Code, fixes with Cursor.
Getting Spun Up: Before writing a feature, uses Claude Code to explore: 'tell me where off happens,' 'tell me similar features.' Builds mental picture, then works with Claude.
Closing: Thank you; will stay to mingle and answer questions.

Transcript

Summarized by YouSumi from the video's transcript. AI summaries can miss nuance — check the video for exact quotes. Report this summary

Summarize your own video

Paste any YouTube link and get a TL;DR, key points and the full transcript like this one — in seconds.

3 free summaries a day·No credit card·6 output languages