Claude Code Field Guide
19 / 35

Skills

Pre-written expert playbooks Claude reads before it builds — install one, or write your own for the tasks you do repeatedly.

Updated since the video
Currency check: Concept fully intact; the format has become considerably richer (dynamic context injection, arguments, subagent execution).

Claude Code is a capable generalist, but for a specific task type — writing a social media strategy, designing a frontend, doing a security review — it does noticeably better with specialized instructions to follow. Skills are those pre-written instructions: when your request matches a skill's description, Claude reads that file (and any supporting reference files it links to) before it starts working. They cost more context than not using one, but the quality gain is usually worth it for anything you do often.

What's grown since the video: skills can now inject live, dynamic context directly into themselves — for example embedding the output of git diff right into the skill's instructions at invocation time, so Claude works from current data rather than static text. They can accept arguments (/fix-issue 123), be restricted to human-only invocation for side-effecting actions like deployment, and run inside an isolated subagent context. Skills are also now the standard way plugins (next topic) distribute reusable functionality.

A copywriting skill for consistent brand voice

A PM installs or writes a copywriting skill with the company's tone-of-voice rules and example headlines, so every marketing-copy request — from any team member using Claude Code — produces on-brand output without re-explaining the style each time.

A PR-review skill with live data

A custom review skill injects the current PR's diff and comments directly via dynamic context syntax, so invoking /review-pr always reviews the actual current state of the PR rather than something the user had to paste in manually.