Inside Anthropic's Daily Claude Code Workflow

The gap between what a company ships and what it actually uses internally tells you something. AI LABS spent time digging through Anthropic's repositories, team posts, and leaked source code to document the specific skills and plugins the Claude Code team runs daily. Some are publicly available. Others are internal tools hidden behind CLI flags. All of them reveal how the people building AI coding assistants think about their own workflows.

The Public Tools

The frontend design plugin exists because AI has an aesthetic problem. Left to their own devices, large language models converge on what AI LABS describes as "general AI aesthetics"—the same generic look that screams machine-generated. Anthropic's solution is a specialized plugin that enhances UI and UX beyond simple prompts. It's available in the official plugin marketplace, installable through a straightforward command, and invocable either manually or automatically when Claude determines it's needed.

The batch skill handles a different kind of convergence problem. When you're migrating code from one library to another or performing similar repetitive operations, you want parallelization. But Claude's standard agent system, which already breaks tasks into smaller parts, wasn't enough. The batch command uses separate work trees—isolated copies of your repository—so each agent operates in its own environment without stepping on the others. As the video explains: "The batch command uses separate work trees, which are isolated copies of your repository, letting each agent work in its own environment."

Once you approve the plan, it spawns agents equal to the units of work, each with dedicated prompts and separate work trees. When they finish, the main agent merges everything into the main branch and can even manage the pull request if configured.

The code simplifier plugin does exactly what its name suggests—removes duplicates and unnecessary files while preserving functionality. It's open source, spawns a single agent, and returns a summary of changes. Anthropic later released a more thorough version called simply "simplify" that spawns three agents and evaluates code across multiple metrics. Both see heavy use internally, suggesting that even teams building AI code generators still spend significant time cleaning up after themselves.

The Internal Skills

What Anthropic keeps private is more interesting than what it releases.

The verify skill exists but isn't mentioned in public tweets. AI LABS found it in leaked source code—an internal tool hidden behind a CLI flag. It verifies code changes by running the app, testing from different angles, and automatically fixing failures. The system prompt isn't in the product code; it's injected directly into team members' environments. This explains why it hasn't been released: verification looks different for every project. Some use Playwright, others use different testing frameworks. But the core principle remains constant—the agent needs a way to confirm its changes work as intended.

Skillify captures entire sessions and turns them into reusable skills. Unlike verify, its system prompt is in the source code, so you can see exactly what it does. The purpose is simple: you brainstorm a workflow with Claude, refine it through conversation, and Skillify converts that into a skill.md file you can invoke later. It identifies repeatable processes, required tools and permissions, agents involved, and everything needed to recreate the workflow. The video notes: "The purpose of Skillify is to record a session where you brainstorm a workflow with Claude, refine the process through multiple conversations, and then save it as a reusable skill."

The tech debt skill runs at session end to find duplicated code. It's not open source because tech debt is project-specific, but the creator of Claude Code mentioned his team uses it routinely. AI LABS reverse-engineered a version that analyzes the codebase with multiple agents, reports redundant duplications, creates shared libraries, and updates components to use centralized features. The session ends with cleanup and minimal debt.

Security and Workflow Integration

The security scan command searches for vulnerabilities across input validation, authentication issues, secret management, injection risks, and endpoint exposure. It incorporates multiple security standards directly into the tool and was released publicly after internal use caught multiple issues. This matters more as AI-generated code produces larger volumes of changes—vulnerabilities can slip into production simply because there's too much to review manually.

For repetitive inner-loop workflows, the team uses slash commands like commit-push-PR, which pushes code and opens pull requests. It's open source and available in the official marketplace, containing the exact workflow the creator described in posts.

The dedupe skill auto-detects duplicate GitHub issues. It's internal, not available publicly, but AI LABS reverse-engineered its behavior from linked issues. It parses input, searches for similar issues using GitHub CLI, applies criteria to avoid false positives, and flags duplicates only when hitting around 70% certainty. Comments explain why issues match, but human verification is still required. For projects with multiple contributors, it prevents wasted time on already-resolved problems.

The Remotion skill creates motion graphics and videos from simple prompts. According to the video, Anthropic's marketing team uses it for all product announcement videos—one of the most important skills in their non-development workflow.

What This Reveals

The distance between public and private tools is instructive. Anthropic releases generic solutions—design enhancement, code simplification, security scanning—while keeping project-specific automation internal. The verify skill stays private because verification looks different everywhere. The tech debt skill stays private because debt accumulates differently in different codebases. Skillify's prompt is public, but each team needs to build their own version tailored to their processes.

This matches what I've observed across fifty years of technology coverage: the tools that matter most are the ones you can't buy. Every company builds internal automation that perfectly fits its workflow, and those tools rarely transfer cleanly to other contexts. What Anthropic has done here is release the frameworks—batch processing, simplification, verification patterns—while keeping the implementations private.

The leaked source code and reverse engineering add an interesting wrinkle. These aren't trade secrets Anthropic is protecting; they're simply tools too specific to be useful elsewhere. But by documenting what they do and how they work, AI LABS has provided a template. You can build your own verify skill, your own tech debt detector, your own duplicate issue finder. You just can't use Anthropic's exact versions, because they wouldn't work for you anyway.

The real question is whether this level of tooling becomes standard or remains the province of companies building AI systems. If you need a dozen specialized skills to make AI coding assistants productive, that's a significant barrier to adoption. If these tools eventually ship as standard features, AI-assisted development looks very different than it does today.

Bob Reynolds is Senior Technology Correspondent for Buzzrag