Skip to main content

Skills

A skill is a structured prompt file that encodes a workflow as a slash command. When you type /brainstorm or /work, you’re invoking a skill — a Markdown document with YAML frontmatter that gets loaded into the agent’s context.

How Skills Work

  1. You type /brainstorm in Claude Code
  2. Claude Code finds skills/brainstorm/SKILL.md
  3. The skill’s content is injected into the conversation
  4. The agent follows the skill’s instructions
Skills are not code — they’re prompt programs. They define phases, decision trees, output formats, and validation steps that the agent follows.

Skill Anatomy

Every skill lives in skills/<name>/SKILL.md with this structure:
---
name: brainstorm
description: Explore ambiguous or early-stage ideas interactively
---

# Brainstorm Skill

## Phase 1: Understand the Idea
...

## Phase 2: Explore Constraints
...
The frontmatter defines:
FieldPurpose
nameSkill identifier — matches the /command name
descriptionOne-line summary shown in skill listings

Built-in Skills

Genie ships with 14 skills organized by function:

Core Pipeline

SkillCommandPurpose
Brainstorm/brainstormExplore ideas interactively, track wish-readiness
Wish/wishConvert a design into a structured WISH.md
Work/workExecute a wish — dispatch agents per execution group
Review/reviewValidate work against criteria — SHIP or FIX-FIRST

Debugging & Fixing

SkillCommandPurpose
Trace/traceInvestigate unknown issues, isolate root cause
Fix/fixApply minimal fix for FIX-FIRST gaps
Report/reportFull bug investigation with GitHub issue creation

Orchestration

SkillCommandPurpose
PM/pmFull PM playbook — triage, assign, track, report, escalate
Council/council10-perspective architectural review
Dream/dreamBatch-execute SHIP-ready wishes overnight
Genie/genieTransform any session into a Genie orchestrator

Utilities

SkillCommandPurpose
Wizard/wizardGuided onboarding — scaffold to first wish
Refine/refineOptimize prompts via prompt-optimizer
Docs/docsAudit and generate documentation
The /learn skill was removed as a standalone skill — behavioral corrections are now handled through Claude Code’s native memory system. The qa-runner agent has been merged into the /qa QA system accessible via genie qa.

Invoking Skills

Skills are invoked as slash commands:
/brainstorm Add a webhook system for real-time notifications
You can pass arguments — the text after the command becomes context for the skill. Some skills are also triggered automatically:
  • /work is invoked by the team-lead when dispatching execution
  • /review is invoked after work completion
  • /fix is invoked when review returns FIX-FIRST

Skill Composition

Skills can delegate to other skills:
  • /wizard delegates to /brainstorm for identity shaping and /wish for plan creation
  • /work invokes /review after execution completes
  • /report cascades through /trace before filing an issue
  • /dream orchestrates /work/review loops across multiple wishes

Creating Custom Skills

To add a custom skill:
  1. Create skills/<name>/SKILL.md
  2. Add YAML frontmatter with name and description
  3. Write the skill’s instructions in Markdown
  4. The skill becomes available as /<name>
---
name: deploy
description: Guide deployment to staging and production environments
---

# Deploy Skill

## Phase 1: Pre-flight Checks
- Verify all tests pass
- Check CI status
- Confirm target environment

## Phase 2: Deploy
...
Skills installed via the plugin system (in ~/.claude/plugins/genie/skills/) are available globally across all projects.

Skill vs. Agent

SkillAgent
WhatA prompt programA running process
LifetimeSingle conversation turnPersistent tmux session
StateStateless (context is the conversation)Stateful (worker registry, mailbox)
Invocation/commandgenie spawn <role>
Use caseStructured workflowLong-running autonomous work
Skills tell agents what to do. Agents are who does it.