Stop Being an AI ‘After‑Hours Engineer’—Let AI Write Better Code

Problem: Re‑educating AI coding assistants

Each new session with an AI coding assistant requires re‑telling the model the project's framework, coding standards, commit‑message format, and deployment workflow, which is repetitive and slows development.

Skill definition and loading

A Skill is a folder containing a SKILL.md file that serves as an operation manual for the AI. Claude reads only the Skill name and description at startup, keeping token usage low. When a task matches a Skill, Claude loads the full Skill content on demand, avoiding unnecessary context consumption.

Open standard support

The Skill format follows the Agent Skills open standard. Implementations that support it include Claude Code, OpenAI Codex CLI, Cursor, Gemini CLI, and GitHub Copilot, allowing a single Skill to be reused across tools.

Skill types

Ability‑enhancing Skills add capabilities the model could not perform before, e.g., operating PDFs, running Playwright browser tests, or scraping web data.

Preference‑coding Skills keep existing capabilities but enforce team‑specific conventions such as a custom commit‑message format, a code‑review checklist, or a weekly‑report template.

Both types can be triggered automatically or manually with the command /skill-name.

Placement

Personal level : ~/.claude/skills/ – available to all projects in the current environment.

Project level : .claude/skills/ – committed to Git so the whole team can use the Skill.

Skill vs. MCP

Model Context Protocol (MCP) provides the pipeline for AI to connect to external tools and data (e.g., databases, APIs). A Skill supplies the static instruction set that tells the AI how to use those tools. Analogy: MCP is the hammer; the Skill is the instruction for hammering a nail.

Ready‑made Skill collections

Curated directory: https://skills.sh

Superpowers – GitHub repository https://github.com/obra/superpowers (200 K stars). Orchestrates multi‑Agent development: brainstorm → plan → split tasks → sub‑Agents → TDD → code review → merge.

GStack – GitHub repository https://github.com/garrytan/gstack. YC‑backed team that turns Claude Code into a full‑stack pipeline with 23 Skills covering architecture, design, code review, QA, and browser testing.

Frontend Design – Anthropic Skill at https://github.com/anthropics/skills/tree/main/skills/frontend-design. Disables overused AI UI styles and forces Claude to consider visual direction before generating code.

Vercel agent‑skills – GitHub repository https://github.com/vercel-labs/agent-skills. Includes:

Web Design Guidelines – 100+ UI rules with file‑line references.

React Best Practices – 57 performance rules, e.g., prioritize fixing request waterfalls before micro‑optimizations like useMemo or React.memo.

Composition Patterns – Refactors Boolean prop overload (e.g., isCompact, showHeader, isRounded) using composition.

Document Skills – Anthropic suite at https://github.com/anthropics/skills/tree/main/skills/pdf. Provides Python‑backed operations for PDF text/table extraction, Word annotation editing, Excel formula handling, and PowerPoint template generation. Individual Skills can be combined (e.g., extract data from PDF → generate Excel report → create PPT).

Webapp Testing – GitHub repository https://github.com/anthropics/skills/tree/main/skills/webapp-testing. Uses Playwright to control a real Chrome window. Example command:

/brainstorm "我想给笔记应用加一个实时协作功能"

causes Claude to ask clarification questions, produce a design document, split the work into 2‑5‑minute sub‑tasks, and execute each in an isolated Agent context.

Trail of Bits Security – GitHub repository https://github.com/trailofbits/skills. Packages professional static‑analysis workflows (CodeQL, Semgrep) used by the security research firm. Requires local installation of CodeQL or Semgrep.

Creating custom Skills

Anthropic provides a Skill Creator at https://github.com/anthropics/skills/tree/main/skills/skill-creator. It runs an interactive Q&A session that asks for the workflow to automate, trigger conditions, required scripts or reference files, and edge‑case handling, then generates a complete Skill directory.