Skills System
The Skills system connects you with a curated marketplace of reusable AI skills. Browse skills matched to your tech stack, install them with a single command, and customize them with learnings that improve over time.
Quick Start
Get recommendations based on your project:
/agileflow:skill:recommendInstall a skill:
npx skills add vercel/next-skills
See what you have installed:
/agileflow:skill:listHow It Works
Marketplace Discovery
AgileFlow includes a curated catalog of ~60 vetted skills across 6 categories:
| Category | Count | Examples |
|---|---|---|
| Frontend | 13 | React, Next.js, Vue, Svelte, Angular, Tailwind |
| Backend | 11 | GraphQL, Express, FastAPI, Node.js, Go |
| Database | 8 | Prisma, Supabase, MongoDB, PostgreSQL, Redis |
| Testing | 10 | TDD, Jest, Playwright, Cypress, pytest, Vitest |
| DevOps | 9 | GitHub Actions, Docker, Kubernetes, Terraform |
| Security | 8 | OWASP, code review, auth patterns, validation |
Smart Recommendations
When you run /agileflow:skill:recommend, AgileFlow:
- Detects your tech stack from
package.json,requirements.txt,go.mod, and project files - Matches against curated skills by scoring tag overlap
- Filters out already-installed skills from
.claude/skills/ - Shows ranked results with install commands
Live Marketplace
For broader discovery beyond curated picks:
npx skills search react # Search by keyword npx skills find # Browse full marketplace
What Are Skills?
A skill is a structured prompt template that teaches Claude how to handle specific tasks in your project context. Skills are:
- Community-built - Created and maintained by the developer community on skills.sh
- Self-improving - Learn from your corrections via the learnings system
- Cookbook-based - Multiple use-case workflows in a single skill
Skill Structure
Each installed skill contains:
.claude/skills/<skill-name>/
├── SKILL.md # Main skill file with instructions and triggers
├── references.md # Links to docs, code patterns
└── cookbook/ # Per-use-case workflows
├── use-case-1.md
└── use-case-2.md
How Skills Work
- Activation: When you ask Claude about a topic matching a skill's triggers, the skill activates
- Execution: Claude reads SKILL.md and follows the instructions
- Learning: When you correct Claude, the skill learns your preferences
- Evolution: Each correction updates the learnings file, making the skill smarter
Installing Skills
From Recommendations
/agileflow:skill:recommendThis shows matching skills with install commands. Copy and run:
npx skills add vercel/next-skills
From Marketplace
Browse skills.sh directly:
npx skills search <keyword> npx skills find
Manual Import
Copy a skill directory to your skills folder:
cp -r /path/to/skill .claude/skills/The Learnings System
Skills can improve automatically through a learnings system:
.agileflow/skills/_learnings/
├── supabase-swift.yaml
├── api-client.yaml
└── react-components.yaml
What Skills Learn
When you use a skill and correct Claude's output, the skill learns:
- Explicit corrections - High confidence learnings from direct feedback
- Approved responses - Medium confidence learnings from approvals
- Patterns - Low confidence observations for review
Learning File Format
Each skill has an optional learnings file at .agileflow/skills/_learnings/<skill>.yaml:
skill: supabase-swift
version: 1
last_updated: 2025-01-08T10:30:00Z
preferences:
code_style: "swift-concurrency"
error_handling: "Result<T, Error>"
conventions:
naming: "camelCase for variables, PascalCase for types"
testing: "Unit tests required for all public APIs"
anti_patterns:
- "Avoid callback patterns (use async/await instead)"
- "Never hardcode API keys in code"
metadata:
corrections_count: 12
approvals_count: 8How Self-Improvement Works
On invocation:
- Claude checks if a learnings file exists
- Reads preferences, conventions, and anti-patterns
- Applies them while executing the skill
On correction:
- You provide feedback or correction
- Claude extracts the signal (what to change)
- Updates the learnings file
- Next invocation uses the updated preferences
Best Practices
Choosing Skills
- Start with recommendations - Run
/agileflow:skill:recommendfirst - Match your stack - Skills work best when they match your actual technologies
- One domain per skill - Focused skills are more effective than broad ones
- Check quality - Review the skill's SKILL.md before using it extensively
Maintaining Skills
- Review learnings - Check what the skill learned from corrections
- Update references - Keep documentation links current
- Archive unused skills - Clean up skills you no longer use
- Share improvements - Commit learnings to git for team consistency
Sharing Skills
Version Control
Commit skills and learnings to git:
git add .claude/skills/
git add .agileflow/skills/_learnings/
git commit -m "Add next-best-practices skill with learnings"Team members automatically get the skill and learnings.
Creating Custom Skills
You can create your own skills manually:
- Create a directory in
.claude/skills/<skill-name>/ - Add a
SKILL.mdwith frontmatter (name,description) - Optionally add
references.mdandcookbook/entries - The skill will appear in
/agileflow:skill:list
Troubleshooting
Skill Not Activating
- Check the "When to Use" section in SKILL.md
- Make sure your question matches the activation triggers
- Edit the skill to add more keywords if needed
Learnings File Missing
- Create the learnings file at
.agileflow/skills/_learnings/<skill>.yaml - Future corrections will be tracked automatically
Related Commands
/skill:list- View installed skills and marketplace recommendations/skill:recommend- Get personalized skill recommendations for your tech stack
On This Page
Skills SystemQuick StartHow It WorksMarketplace DiscoverySmart RecommendationsLive MarketplaceWhat Are Skills?Skill StructureHow Skills WorkInstalling SkillsFrom RecommendationsFrom MarketplaceManual ImportThe Learnings SystemWhat Skills LearnLearning File FormatHow Self-Improvement WorksBest PracticesChoosing SkillsMaintaining SkillsSharing SkillsVersion ControlCreating Custom SkillsTroubleshootingSkill Not ActivatingLearnings File MissingRelated Commands