Claude Certified Architect Study Guide

All skills include a name and a description in their frontmatter. Claude loads the names and descriptions of all skills at the start of each session. When you ask it to do something, it will use the description to determine if a skill needs loaded into context. If it deems a skill relevant, then (and only then) will it read the contents of the full skill.

Personal skills go in ~/.claude/skills (your home directory). These skills will be used across all of your projects. Skills saved in a project/repo root (.claude/skills) will be shared with anyone who clones the repo.

Put information in CLAUDE.md that you want Claude to always have in context. If there are details you want it to know/use that are task-specific, that’s what a skill is for.

Claude loads skill descriptions at the start of each session, so simply restart Claude.

Avoid conflicts by changing the skill names and descriptions so they are unique.

If two skills share the same name, enterprise skills will be used first, then personal skills, then project skills, then plugins.

You can find out more on the open standard spec. Name and description are the only two required. Additional fields include:

• license to reference a bundled license for the skill
• compatibility can indicate intended product, required system packages, etc.
• metadata accepts key-value pairs to store additional properties not in the spec
• allowed_tools (still experimental) defines tools that are pre-approved to run.

Skills should be 500 lines or less. This helps keep the needed context window small. The goal is to have the skill act as a table of contents, not the entire book. If you need to put more information in, use script, reference, or asset files that Claude can load in if necessary:

• Put executable code in scripts/
• Add documentation to references/
• Put images, templates, and other data files in assets/

Files in your scripts/ directory should be actual executable code. In your SKILL.md file, if you tell Claude to run a script (not read it), then it will run the script and only the output will get loaded into context.

The goal is for Claude to load a skill when you want it to. A good description describes what a skill does and tells Claude when to use it. If it isn’t being loaded when you expect it to, add keywords you typically use in prompts to the description.

The default Claude subagents NEVER use custom skills. By default custom subagents you define WON’T EITHER. To give a subagent access to a skill, explicitly list the skill in the frontmatter of the subagent’s .md file. This skill will be loaded into context EVERY TIME the subagent is used.

Project-specific skills can be version controlled in the repository. You can also defiine plugins that can be shared in a marketplace.

1. Try the agent skills verifier command (skills-ref) to ensure the skill is valid.
2. If it doesn’t trigger, update the description so it overlaps with how you phrase your requests.
3. If skills aren’t loading, they need to be in the right structure and in the right location.
4. Run Claude in debug mode claude --debug and try loading the skill.
5. If scripts can’t be run, make sure chmod +x is applied to all scripts.

Answer

Answer

Answer