How Anthropic Employees Effectively Use Claude Skills to Enhance Workflow and Output Quality

Anthropic, the inventor of Claude, has recently released their internal playbook on how they use Claude skills across various teams including engineering, marketing, and legal. These skills have helped streamline both technical and non-technical workflows. After analyzing interviews, blog posts, and official documentation from Anthropic, this article distills their approach into five actionable lessons anyone can implement today.

---

Lesson 1: Understand the Categories of Claude Skills

Anthropic categorizes Claude skills into nine technical categories, which can be grouped into four main skill types applicable to both technical and non-technical users:

1. Utility Skills

These are small, reusable skills designed to perform one specific task. They are often combined with other larger skills.

• Technical example: Interacting with specific libraries or APIs.
• Non-technical example: Drafting a response in a specific voice or simplifying writing.

2. Verification Skills

Verification skills check the final output quality and correctness. Anthropic notes these have the most measurable impact on Claude's output quality internally.

• Technical example: Product verification or code quality review.
• Non-technical example: Brand voice check or style guide review.

3. Data Enrichment Skills

These skills pull external data into the system to enhance the final product.

• Technical example: Pulling website performance or conversion traffic data.
• Non-technical example: Analyzing competitors or pulling recent consumer reports.

4. Orchestration Skills

Orchestration skills chain multiple steps and skills together to create a final output.

• Example: A "generate report" skill that pulls data (data enrichment), drafts a write-up (utility), and verifies branding and numbers (verification).

Important: The best skills fit cleanly into one category. Skills that try to do too much across categories can confuse the agent.

---

Lesson 2: Use the Power Components When Building Skills

A Claude skill is not just a markdown file but a folder containing multiple components such as scripts, assets, and data. The three most important components to optimize are:

Scripts

Scripts are computer code that runs to complete deterministic tasks. They allow Claude to focus on composition and decision-making rather than reconstructing boilerplate code.

• Deterministic workflows produce the same output for the same input (e.g., 2 + 2 = 4).
• Non-deterministic workflows may produce different outputs for the same input.

Using scripts for deterministic parts reduces token usage and increases repeatability.

Assets/Templates

You can upload specific files as templates (e.g., PowerPoint presentations) in the assets folder. Claude uses these templates as starting points for outputs, ensuring consistency and easy updates.

Setup Prompts

Clear and user-friendly setup prompts are essential for skill usability. Key practices include:

• Creating a config.json file to store setup values and remember them for future runs.
• Using the "ask user question" tool to present structured multiple-choice options.
• Declaring an arguments field at the top of the skill markdown file to remind users of required inputs.

Building skills with future usability in mind ensures sustainability and ease of maintenance.

---

Lesson 3: Focus on Verification to Amplify Output Quality

Verification skills are the most impactful for improving Claude's output quality. Anthropic recommends dedicating time to perfect these skills.

Two Types of Verification:

1. Verify for Correctness: Ensures facts, numbers, or app functionality are accurate.
2. Verify for Quality: Ensures the output meets the desired quality standards.

Verification acts as a quality amplifier, raising the bar on every output.

Setting Up Verification Skills:

• Anthropic provides built-in skills like /verify to run and confirm code changes and /run to launch apps for self-review.
• Custom verification skills should produce objective outputs (e.g., pass/fail, grades).
• Skill-driven verification involves adding verification components to existing skills for easier quality checks.

Example:

Amol Agrawal, Anthropic's head of growth, created a verification skill simulating feedback from his manager by feeding Claude data from her public writings and Slack conversations. This skill provides weekly feedback, effectively acting as an AI clone of his manager's judgment.

Similarly, the author uses an "internal focus group" skill to get expert feedback on projects, which inspired the creation of buildpartner.ai, a platform offering expert advice via Claude skills.

---

Lesson 4: Write the Gotchas to Build a Personal Moat

Gotchas are a running list of issues Claude encounters when using a specific skill. They serve as a list of things not to do within the skill markdown file.

Importance of Gotchas:

• Gotchas document common failure points and edge cases.
• They evolve over time as more issues are discovered.
• Adding gotchas encourages continuous testing and improvement.

Best Practices:

• Start with a few lines and a single gotcha; improve as you identify new edge cases.
• Only include gotchas based on actual failures, not hypothetical ones.
• Format gotchas in a way that allows easy approval or denial of changes (e.g., markdown files for Obsidian).

Gotchas act as a personal moat, protecting the quality and reliability of your skills.

---

Lesson 5: Tune the Trigger to Control Skill Invocation

A trigger is what invokes a skill. For example, typing /draft email manually triggers the draft email skill.

Automatic Triggering:

Claude automatically decides when to use a skill by scanning a list of available skills and their descriptions at the start of a session.

Writing Effective Descriptions:

• The description should specify when to trigger the skill, not just summarize it.
• It should name who the skill serves and under what conditions it should fire.

Example:

The "front-end design" skill description includes phrases like "Use the skill when the user asks to build web components, pages, or applications," enabling Claude to trigger it automatically when such requests are made.

Orchestration Skills Calling Other Skills:

• Orchestration skills can reference and invoke other installed skills by name.
• Complex workflows should be broken into sub-utility skills to avoid overlap and improve maintainability.
• Updating a utility skill automatically updates all orchestration skills that call it.

---

Conclusion

By understanding the categories of Claude skills, leveraging power components, focusing on verification, documenting gotchas, and tuning triggers, Anthropic employees have created a robust system to enhance workflow efficiency and output quality. These five lessons, battle-tested by Anthropic and the author, provide a comprehensive framework for anyone looking to optimize their use of Claude skills.

Implementing these strategies will help you build sustainable, maintainable, and high-quality AI-powered workflows.

---

If you found this guide helpful, consider exploring further insights on how Claude's founders determine what is worth building to maximize the potential of AI skills.