CHAPTERS
- 0:00 – 0:31
Claude Code comes to the web: browser-based, cloud-run coding sessions
The video opens by announcing Claude Code’s availability on the web and mobile, removing the need to start in a terminal or IDE. It explains the basic workflow—connect GitHub, choose a repo, describe the task—and highlights that execution happens in Anthropic’s cloud.
- •Claude Code is now usable directly from a browser or the Claude mobile app
- •Connect a GitHub account, select a repository, and delegate a coding task via prompt
- •Work runs on cloud infrastructure rather than your local machine
- •Designed to make quick changes without opening terminal/IDE
- 0:31 – 1:02
Parallel work across repos: tackling backlogs, fixes, and new features simultaneously
The host emphasizes the ability to run multiple Claude Code tasks at the same time, even across different repositories. This is positioned as a practical way to clear bug backlogs, handle routine maintenance, and prototype new features in parallel.
- •Multiple sessions can run concurrently
- •Parallel tasks can span multiple repositories
- •Useful for bug backlogs and routine fixes
- •Also suited to building new ideas/features quickly
- 1:02 – 1:32
Web demo setup: prompting a real repo change (registration validation + password strength)
A live demo shows the web UI at claude.ai/code, selecting a repository, and issuing a concrete prompt. The requested work focuses on input validation and enforcing strong passwords across both frontend and backend.
- •Navigate to claude.ai/code and select a repo
- •Prompt example: add input validation to user registration endpoint
- •Ensure strong password rules on both frontend and backend API
- •Demonstrates quick task delegation from the web
- 1:32 – 2:02
How the web agent executes: isolated sandbox, Git access, real-time progress
While the task runs, the host describes the execution environment. Claude operates in an isolated sandbox with full Git access, and the user can watch the work progress in real time.
- •Claude runs in an isolated sandbox environment
- •Full Git access is available during execution
- •Progress is visible live while the agent works
- •Supports a “delegate and monitor” workflow
- 2:02 – 2:32
Claude Haiku 4.5 announcement: fast, cost-efficient coding performance
The video shifts to model updates: after Sonnet 4.5’s release, Haiku 4.5 arrives as a faster, cheaper model with surprisingly strong coding capabilities. It’s framed as delivering near–state-of-the-art performance at a fraction of the cost.
- •Sonnet 4.5 positioned as top-tier coding model; Haiku 4.5 released shortly after
- •Haiku 4.5 offers similar coding performance to earlier SOTA at ~1/3 cost and >2× speed
- •Surpasses Sonnet 4 on some tasks (e.g., computer use)
- •Highlights rapid iteration in model capability over months
- 2:32 – 3:03
Benchmarks + pricing: where Haiku 4.5 lands versus Sonnet
Concrete metrics and economics are provided to help developers gauge fit. The host cites benchmark scores and token pricing, emphasizing Haiku’s strong agentic coding performance relative to Sonnet 4.5.
- •Benchmark callouts: 73.3% Suitebench Verified, 41% TerminalBench
- •Achieves ~90% of Sonnet 4.5 performance in agentic coding evals
- •Availability: “wherever you get your Claude”
- •Pricing: $1/M input tokens and $5/M output tokens
- 3:03 – 3:33
When to use Haiku 4.5: multi-agent workflow (Sonnet plans, Haiku implements)
The recommended usage pattern is a division of labor between models. Sonnet 4.5 is used for high-level decomposition and planning, while Haiku 4.5 is used to implement the code quickly and cheaply—especially for large refactors and tech debt.
- •Common pattern: Sonnet 4.5 for planning/problem decomposition
- •Haiku 4.5 for implementation and code-writing
- •Multi-agent approaches gaining popularity
- •Particularly effective for large-scale refactors and technical debt work
- 3:33 – 4:03
Checking results in Claude Code web: change summaries across backend and frontend
The host returns to the running web task and reviews the output. Claude provides a summary of changes spanning backend (e.g., service.go) and frontend TypeScript updates.
- •Claude produces a clear summary of code changes
- •Edits span backend files (service.go)
- •Edits span frontend TypeScript as well
- •Demonstrates end-to-end task completion visibility
- 4:03 – 4:33
Seamless handoff to local dev: “Open in CLI” preserves context and progress
A highlighted feature is the “Open in CLI” button, which transfers the web session into a local terminal environment. This preserves chat context and changes, enabling continued work in tools like VS Code without losing state.
- •“Open in CLI” generates a command to resume locally
- •Local environment inherits web context and progress
- •See messages and changed files after handoff
- •Continue in VS Code and switch between phone/browser/local fluidly
- 4:33 – 5:05
Broad prompts made safe: interactive follow-up questions for correct implementation
The final product feature shown is Claude Code’s ability to ask clarifying questions when prompts are underspecified (e.g., “add caching”). The system collects structured choices (layer, goals, target area, tech) before implementing.
- •Claude asks follow-ups when the prompt is too broad
- •Caching example: choose level (DB query vs API response caching, etc.)
- •Specify goals (performance vs reducing DB load) and target areas (user/workout/analytics data)
- •Choose technology (Redis, in-memory, Postgres) before implementation
- 5:05 – 5:14
Wrap-up: web availability, Haiku 4.5 positioning, and call to action
The video closes by summarizing the main updates: Claude Code on the web for parallel cloud development and Haiku 4.5 for fast, low-cost coding. It ends with availability details and an invitation for Pro/Max users to try it.
- •Claude Code on the web enables parallel cloud-based development
- •Haiku 4.5 positioned as lightning-fast and cost-effective
- •Mentions “smarter local workflows” enabled by web→CLI continuity and follow-up questions
- •Availability: Claude Code web for Pro and Max users at claude.ai/code
