Insights on Claude Code (from Latent Space Podcast)

09 May, 2025

On May 8th, Latent Space Podcast had a great episode on Claude Code featuring Catherine Wu and Boris Cherny. The podcast is packed with amazing insights on the design and philosophy behind Claude Code.

Sharing my notes on what I learned.

Video
Transcript
Claude Code changelog

CC = Claude Code

Anecdotes and Concrete Use Cases

• CC is writing ~80% of its own code. But humans still review everything.
• The night before launch, Boris couldn't find a good markdown parser, so he had CC write one from scratch. It's still used in production.
• In the internal GitHub repo, they use a GitHub Action that invokes CC in non-interactive mode to do intelligent linting. It checks that the code matches the comments, makes changes, and commits back to the PR via GitHub MCP.
• Boris: "Megan the designer on our team, she is not a coder but she's writing pull requests. She uses code to do it. She designs the UI. Yeah. And she's landing PRs to our console product."
• When considering a feature, Boris has CC prototype multiple versions to help him decide.
• Boris builds UIs by giving CC a screenshot and iterating with Puppeteer MCP until the result matches the mock.

Productivity Boosts from CC

• Boris personally reports a 2× productivity boost from CC.
• Some engineers see a 10× boost, others only 10%. It depends on how well they adapt to CC.
• Cat: "Sometimes we're in meetings and sales or compliance says 'we need X feature,' and 10 minutes later Boris says, 'All right, it's built. I'm going to merge it later. Anything else?'"
• Bugs reported by support are often fixed by CC within 10 minutes.
• CC enables engineers to build features that would otherwise stay in the backlog.
• Anthropic is working on measuring customer productivity gains.

How CC Got Started

• Inspired by the open-source tool Aider. Anthropic had an internal tool called Clyde, slow but capable of writing PRs. It made Boris "AGI-pilled."
• CC began as a research experiment. Boris wired Claude into a terminal and got it to write usable code.
• Early on they saw very high adoption inside Anthropic. This led to giving it a dedicated team.
• Like Artifacts and MCP, CC started bottom-up, driven by developers building internal tools.

Product Philosophy

• Do the Simple Thing First: The simplest implementation is the often best. For example, to add memory, they considered vector stores, but just had CC read/write CLAUDE.md markdown files.
• Keep teams small and operations scrappy. Scale only when you see PMF.
• Heavy internal dogfooding. CC became popular with engineers and researchers internally.
• Product managers are lightweight. Engineers drive product decisions.
• Instead of writing docs, they prototype with CC and test immediately.
• Roadmap is shaped by anticipated model capabilities, always looking ~3 months ahead.
• The team rewrites CC every 3–4 weeks for simplicity and optimization.

Comparison with Cursor, Windsurf, etc.

• Cursor/Windsurf have PMF today. CC is more experimental, aiming at early adopters.
• CC is a thin wrapper over the model. Scaffolding is minimal, "bitter lesson" philosophy.
• Designed for power users. Offers raw model access.
• Supports parallel workflows (e.g. "fix 1,000 lint violations at once").
• Optimizes for capability, not cost.

Open Source

• CC is not open source, but they're "investigating."
• OS would be high maintenance for them.
• No secret sauce: CC is a thin JavaScript wrapper, and people have already decompiled it.

Cost

• Originally pay-as-you-go based on token use.
• Now part of Anthropic's Max plan.
• CC prioritizes smart capabilities over being the cheapest tool.
• Free for internal users, some are spending thousands of dollars a day with it.
• Cat: "Currently we're seeing costs around $6 per day per active user."
• Boris: "It's an ROI question, not a cost question... Engineers are expensive, and a 50–70% productivity gain is worth it."

UI / UX

• Boris: "It's really hard to design for a terminal. There's not a lot of modern literature on it."
• Making a terminal app intuitive and fresh took real design work.
• Inconsistent terminal behavior feels like early web design, "IE6 vs Firefox."
• CC wrote its own markdown parser.
• Technologies used: React Ink (React → terminal escape codes), Commander.js, Bun for bundling.
• Anthropic has a dedicated designer and is creating a terminal-first design language.

Interactive vs. Non-Interactive Mode

• Interactive mode: Default. You approve CC's actions. Good for trust-buildin, complex tasks and asking questions.
• Non-interactive mode (-p): CC runs end-to-end without user intervention. Good for automation, CI/CD, batch ops.
• Used for massive-scale tasks like analyzing/updating thousands of test cases.
• Works like a Unix tool, composable. A finance user pipes CSVs into CC to query them.
• Less secure/predictable. Should be used for read-only tasks unless well controlled.

Memory and Context

• Memory = simple claude.md markdown file loaded into context.
• Auto-compact simulates "infinite context" by summarizing past messages.
• Users can send # to create memory entries.
• Early prototypes used RAG and vector stores but switched to agentic search (e.g. grep, glob) which performs a lot better based on benchmarks and vibes.
• RAG issues: complexity in indexing (how to store the index and keep it in sync); external dependencies -> security concerns. Agentic search sidesteps these issues at the cost of latency and tokens.
• No between-session memory yet. They want to support cases where users want a fresh start vs. resuming with full history, similar to git branches.
• Bitter lesson: eventually the model will manage memory on its own.

Custom Slash Commands

• Users can create local, reusable /commands tied to markdown prompt files.
• These files accept CLI arguments.
• Example: /lint command linked to a list of linting rules.
• Unlike MCP, slash commands are just prompts, not tools.

MCP Integration

• CC acts as both MCP client and (eventually) server.
• As client: CC uses tools like Puppeteer, GitHub API via MCP.
• As server: could expose local workflows to be used by AI clients, though this isn't implemented yet.

Changes Since Podcast Recording

• CC is now included in the Max plan no extra cost, removes cost anxiety.
• CC now supports web search.