Skills and Rules

Skills and rules are reusable workflow units. Instead of repeating the same instructions every time you ask for a code review, a migration, or a deployment task, you package that knowledge once and let the agent activate it when needed.

Quick Take

Use a context file for always-on rules. Use a skill when the workflow is specialized enough that you do not want it in every session.

When To Use Them

Use skills or rule bundles when:

• a task type repeats across projects
• the workflow has multiple steps
• the agent needs access to supporting references, scripts, or templates
• you want consistency without turning your base context file into a wall of text

When Not To Use Them

Do not turn every tiny preference into a skill. If the rule is universal for the repo, put it in the main context file. If the instruction is only for one task, keep it in the prompt.

A Simple Split

Put it in…	If the instruction is…
Context file	stable, repo-wide, and almost always relevant
Skill or rule bundle	specialized, reusable, and task-triggered
Task prompt	specific to the change you want right now

Skills Vs Context Files

Question	Context file	Skill
When loaded	every session	only when relevant
Best use	repo-wide defaults	repeatable workflows
Scope	always-on	task-scoped
Good examples	commands, conventions, safety	deploy, review, migration

Examples

• A code-review skill that tells the agent to look for bugs, regressions, and missing tests first.
• A release-check rule that always verifies changelog, versioning, and smoke tests.
• A frontend-qa workflow that asks for screenshots and accessibility checks after UI edits.

Example Skill Layout

.claude/
  skills/
    deploy/
      SKILL.md
      checklist.md
    pr-review/
      SKILL.md

Starter SKILL.md

---
name: deploy
description: Run the project deployment workflow safely
---

# Deploy Checklist

1. Run `pnpm test`
2. Run `pnpm build`
3. Confirm environment changes
4. Deploy using the approved command
5. Smoke test the critical path

Cross-Agent Notes

• Gemini CLI explicitly supports packaged agent skills.
• Codex environments often use local skills or instruction bundles for specialized workflows.
• Claude Code may express the same idea through project memory, commands, or hooks rather than a single “skills” abstraction.

The exact mechanism changes, but the core idea is the same: keep reusable expertise reusable.