CHAPTERS
The problem: repeating the same guidance to Claude
The speaker highlights the friction of repeatedly teaching Claude the same team conventions—coding standards, PR feedback structure, and commit message formats. This sets up the need for a reusable mechanism so that preferences don’t have to be restated every time.
- •Repeatedly explaining coding standards and conventions wastes time
- •PR reviews often require re-stating how feedback should be structured
- •Commit message and formatting preferences keep getting reintroduced
- •Motivation for a reusable, persistent approach
What a “skill” is: teach once, apply automatically
A skill is introduced as a Markdown file that encodes how Claude should perform a task, so the behavior can be reused automatically whenever relevant. The concept extends to “agent skills,” which can include additional assets beyond instructions.
- •A skill is a Markdown file that teaches Claude how to do something once
- •Claude reuses the skill automatically when relevant
- •Agent skills can be folders containing instructions, scripts, and resources
- •Goal: more accurate and efficient agent behavior
How Claude Code selects skills: description-based matching
The transcript explains that Claude decides to use a skill based on its description. When you make a request (e.g., review a PR), Claude compares the request to available skill descriptions and activates the matching ones.
- •Claude Code uses a skill’s description to determine applicability
- •Requests are matched against all available skill descriptions
- •Matching skills are activated automatically for the task
- •Example: a PR review request triggers a PR review skill
Where to store skills: personal vs. project scope
Skills can be stored in different locations depending on who should benefit. Personal skills live in a user directory and follow you across projects, while project skills live in the repository so the whole team gets them when cloning.
- •Personal skills location: ~/.claude/skills
- •Personal skills capture individual preferences (explanations, formats, styles)
- •Project skills location: repo-root .claude/skills
- •Project skills distribute team standards automatically to collaborators
Examples of project skills: encoding team and brand standards
The speaker gives examples of standards well-suited for project-level skills, such as brand guidelines and design system constraints. This emphasizes that skills can represent non-coding conventions too.
- •Team standards belong in project skills
- •Examples: company brand guidelines
- •Design constraints like preferred fonts and colors
- •Ensures consistency across contributors and tasks
Skills vs. Claude.md: on-demand activation vs. always-on context
Claude Code offers multiple customization paths, and skills are distinguished by being automatic and task-specific. Unlike Claude.md, which loads into every conversation, skills only load when needed based on relevance.
- •Claude.md content loads into every conversation
- •Use Claude.md for universal rules (e.g., always use TypeScript strict mode)
- •Skills are task-specific and activate only when relevant
- •Skills reduce unnecessary context usage compared to always-loaded instructions
Context efficiency: skills load minimal metadata until needed
Skills are designed to avoid bloating the context window by loading only their name and description initially. The full checklist or detailed instructions only come into play when the task calls for it.
- •Only the skill’s name and description are considered initially
- •Prevents filling the context window with irrelevant checklists
- •Example: PR review checklist shouldn’t be present during debugging
- •Detailed guidance is applied when the matching task occurs
Skills vs. slash commands: automatic recognition vs. manual invocation
The speaker contrasts skills with slash commands: slash commands require explicit typing, while skills are applied automatically when Claude recognizes the situation. This makes skills smoother for recurring workflows.
- •Slash commands require you to remember and type them
- •Skills apply automatically based on task recognition
- •Less friction for recurring patterns and standards
- •Better fit for “set it and forget it” conventions
Best use cases: specialized, repeatable standards
Skills are positioned as ideal for knowledge that’s specialized and repeatedly needed in specific contexts. Examples include code review standards, commit message formats, and organizational brand rules.
- •Best for specialized knowledge tied to specific tasks
- •Great for code review standards and checklists
- •Useful for preferred commit message formats
- •Applies well to organizational brand guidelines
Rule of thumb: if you repeat it, it should be a skill
The video closes with a practical heuristic for deciding what to encode as a skill. Repeated explanations to Claude are treated as signals that a skill should be written to capture that guidance permanently.
- •Repeated instructions are a strong signal for a skill
- •Skills formalize preferences and standards into reusable assets
- •Reduces repetitive prompting over time
- •Encourages building a library of task-specific guidance
