An open, local-first library of AI agents, skills, prompts, and routines, with a native macOS app to browse, install, and export them.
π€ Contribute a resource Β· π¬ Submission guide Β· π§© Templates Β· π¬ Watch the explainer Β· π» Install the app Β· ποΈ Architecture
βΆ Watch the explainer (16s, with voiceover) Β· made with HyperFrames and ElevenLabs. Source: video/explainer/
Tip
Got a prompt, skill, or agent you use every day? Share it. Copy a template, open a pull request, and once it's merged and the release is approved it reaches every Skilldeck app automatically. It even works entirely in the browser, with no tools to install. Start here β
Skilldeck is two things that live in one repository:
- A community content platform. Agents, skills, prompts, and routines are plain Markdown + YAML packages under
resources/. Every pull request is validated (schema, structure, safety, duplicate detection, compatibility claims). Every merge tomainbuilds a versioned, checksummed catalog, which is published as a GitHub Release once a maintainer approves it. - A native macOS app. Skilldeck ships with the whole catalog bundled, so it works fully offline from first launch. It lets you discover, search, and read resources, then install them into Claude Code, OpenAI Codex, Claude Desktop, the Agent Skills format, or any folder you choose, with a preview of every file it will write.
| Type | Count | What it is | Installs as |
|---|---|---|---|
| Agents | 100 | Complete task-oriented specialists with a system prompt, procedure, output format, and quality bar | Claude Code subagents (~/.claude/agents/) |
| Skills | 100 | Reusable capabilities in the open Agent Skills SKILL.md format |
Claude Code / Codex skills, Claude Desktop skill zips |
| Prompts | 30 | Copy-ready prompts with documented {{PLACEHOLDERS}} |
Claude Code commands, Codex custom prompts |
| Routines | 20 | Repeatable human-guided workflows with steps and checklists | Claude Code commands |
That's 250 resources across 10 domains: software development, AI & machine learning, product & design, data & analytics, marketing & growth, research & intelligence, finance & banking, security & compliance, operations & productivity, and strategy & leadership. They're grouped into 12 curated collections such as RAG Development Toolkit and Complete Software Development Toolkit.
- Discover: featured resources, shelves by goal, most-referenced resources, and curated collections.
- Search across names, descriptions, tags, use cases, and content, with suggestions, filters (type, domain, use case, platform, compatibility level, author, installed state, featured), sorting, and grid/list views. Press βK to search.
- Detail pages rendered from the package Markdown: purpose, inputs, procedure, outputs, examples, limitations, safety notes, related resources, and a prompt panel that fills in placeholders before you copy.
- Install with a dry-run plan that shows every destination file. Existing files are never overwritten silently: you choose cancel, or back up and replace. You can install for your user, a project folder, or a custom directory, and every file is hash-checked after writing.
- Installed view that detects local edits, updates resources when the catalog has a newer version, and uninstalls to the Trash. Files you've modified are kept.
- Export a single package as a folder or zip, a Claude Desktop skill zip, an
AGENTS.mdsection for Codex, or the complete catalog. - Favorites, recents, and personal collections, stored locally in JSON.
- Contributors: profile pages with GitHub avatar, bio, and featured work, cached for 24h and resilient to rate limits.
- Catalog updates: checks for new community resources at launch and periodically (every 6 hours by default) against the rolling
catalog-latestrelease, installs them automatically (optional), highlights them under New in the Catalog, verifies SHA-256 for the bundle and every package, shows a diff of added, updated, and removed resources before applying, and lets you roll back to the previous version or the bundled catalog in one click. - Deep links such as
skilldeck://resource/threat-modeling-agent,skilldeck://collection/rag-development-toolkit, andskilldeck://search/postmortem.
From source (macOS 15+, Xcode 16+):
brew install xcodegengit clone https://github.com/seyhunak/skilldeck.git && cd skilldeck/app && xcodegen generatexcodebuild -project Skilldeck.xcodeproj -scheme Skilldeck -configuration Release -derivedDataPath build/DerivedData buildopen build/DerivedData/Build/Products/Release/Skilldeck.appPrebuilt: every run of the macOS app workflow uploads a Skilldeck-app zip. Builds are ad-hoc signed and not notarized, so on first launch right-click the app and choose Open.
Without the app: every resource is plain files. Copy a skill folder into ~/.claude/skills/, or download packages/<id>.zip from the latest catalog release.
Skilldeck grows through its community. Anyone with a GitHub account can add an agent, skill, prompt, or routine. Every contributor gets a profile page in the app with their GitHub avatar and everything they've shared.
1. You 2. CI (automatic) 3. Maintainer 4. Release (automatic) 5. Every Skilldeck app
βββββββββββββ ββββββββββββββββββββββ ββββββββββββββ βββββββββββββββββββββββ βββββββββββββββββββββββ
copy a template βΆ Validate submission βΆ review & merge βΆ build catalog-v1.N.0 βΆ checks at launch + every
fill it in schema Β· structure Β· into main maintainer approves 6h, verifies SHA-256,
open a PR safety Β· duplicates Β· β latest.json shows "New in the Catalog"
compatibility
main is protected, so a PR can only merge once Validate catalog changes passes. The catalog the app reads is published only after the repository owner or an authorized maintainer approves the release.
In the browser (nothing to install):
- Fork this repo.
- Open the template for your type and copy its two files.
- Add file β Create new file at
resources/<domain>/<use-case>/<your-id>/manifest.yaml(typing/creates the folders), paste, and fill in. Then add the entrypoint file next to it. - Open a pull request. The check's summary lists anything that needs fixing.
On your machine:
gh repo fork seyhunak/skilldeck --clone && cd skilldeck && python3 -m pip install -r requirements.txtcp -R templates/skill resources/software-development/code-review/my-skill-idpython3 scripts/validate_catalog.py --only my-skill-idgit checkout -b add-my-skill-id && git add resources && git commit -m "Add my-skill-id skill" && gh pr create --fill| Type | Template | Entrypoint | Required ## sections |
Min. words |
|---|---|---|---|---|
| Agent: a specialist for a whole task (ships as a Claude Code subagent) | templates/agent/ |
agent.md |
Purpose Β· When to use Β· Required inputs Β· Operating procedure Β· Expected output Β· Supporting skills Β· Example prompts Β· Error handling Β· Limitations Β· Safety boundaries | 650 |
| Skill: a focused capability an agent loads when relevant (Agent Skills) | templates/skill/ |
SKILL.md |
When to use Β· Inputs Β· Procedure Β· Expected output Β· Constraints | 400 |
Prompt: one copy-ready prompt with {{VARIABLES}} |
templates/prompt/ |
prompt.md |
Intended task Β· Required context Β· Prompt Β· Variables Β· Output format Β· Example usage Β· Customization | 300 |
| Routine: a repeatable human-guided workflow | templates/routine/ |
routine.md |
Objective Β· Trigger Β· Prerequisites Β· Inputs Β· Steps Β· Decision points Β· Expected output Β· Validation checklist Β· Human approval | 400 |
| Contributor profile (optional) | templates/contributor/ |
contributors/<login>/profile.yaml |
β | β |
Domains and use cases come from catalog/domains.yaml. The folder path must be resources/<domain>/<first use case>/<id>/.
π Example: a complete skill package (click to expand)
resources/software-development/code-review/sql-migration-review/manifest.yaml
schema_version: "1.1"
id: sql-migration-review # same as the folder name
name: SQL Migration Review
type: skill
description: Reviews a SQL schema migration for locking, data loss, and rollback risk before it runs in production.
purpose: >
Catch the migrations that take a table lock for minutes or silently drop data,
while they are still a pull request.
version: 1.0.0
author:
name: Your Name
github: your-github-login # becomes your profile page + avatar
license: MIT
domain: software-development
use_cases: [code-review]
tags: [sql, migrations, databases]
created: "2026-11-09"
updated: "2026-11-09"
content:
entrypoint: SKILL.md
files: [] # list any extra files, e.g. references/checklist.md
compatibility: # keep within catalog/compatibility.yaml
- {platform: claude-code, level: native_install, resource_types: [skill]}
- {platform: codex, level: native_install, resource_types: [skill]}
- {platform: claude-desktop, level: native_import, resource_types: [skill]}
- {platform: agent-skills, level: file_export, resource_types: [skill]}
- {platform: generic, level: generic_export, resource_types: [skill]}
dependencies: []
related: []
safety:
contains_scripts: false
network_access: false
handles_sensitive_data: false
notes: Reviews migration text only; never connects to a database.
status: submitted # leave as-is; merge + approved release publishes it
tested: false
tested_with: []
changelog: ["1.0.0: Initial submission."]resources/software-development/code-review/sql-migration-review/SKILL.md
---
name: sql-migration-review
description: Review a SQL schema migration for lock duration, data loss, and rollback safety. Use when a pull request adds or changes a migration file.
---
# SQL Migration Review
## When to use
A PR adds a migration (ALTER/CREATE/DROP, backfills). Not for query performance tuning.
## Inputs
- The migration file(s) and the target database engine and version.
- Optional: approximate row counts of the affected tables.
## Procedure
1. Classify each statement: additive, destructive, or rewriting.
2. Flag statements that take an ACCESS EXCLUSIVE lock on large tables ...
3. ... (concrete, checkable steps with thresholds)
## Expected output
A table: statement Β· risk (blocker/major/minor) Β· why Β· safer alternative.
## Constraints
Never suggest running the migration. State the assumed engine version.Real skills are longer (400+ words) with worked examples. Browse resources/ for 250 of them.
- β Concrete: real steps, commands, criteria, and thresholds, not "follow best practices"
- β A defined output someone can act on, with an example
- β Honest limitations and safety boundaries (finance, legal, and security content needs the disclaimer)
- β Original: not a near-copy of an existing resource (the validator flags β₯ 45% overlap)
- β No secrets, scripts, binaries, or symlinks. Only
.md .txt .yaml .yml .json .csvfiles.
π Guides: Submission guide (step by step, plus common validation errors) Β· Contributor guide Β· Package format Β· Compatibility Β· Review policy
π‘ Not ready to write one? Request a resource or report a problem with one.
Improving the app or tooling? See Development & testing. Run cd app && swift test and python3 -m unittest discover -s tests before opening a PR.
resources/<domain>/<use-case>/<id>/ 250 packages: manifest.yaml + agent.md | SKILL.md | prompt.md | routine.md
catalog/ domains, collections, compatibility registry, launch requirements
contributors/<login>/profile.yaml contributor profiles
schemas/ JSON Schemas for manifests, profiles, collections
templates/ starting points for each resource type
scripts/ validate, build, launch gate, quality report (Python, no network)
tests/ tooling tests (unittest)
app/ macOS app: SkilldeckCore Swift package + SwiftUI app (XcodeGen)
.github/workflows/ PR validation, catalog build, catalog release, macOS app CI
docs/ architecture, format, compatibility, security, release process
- Architecture: how the catalog pipeline and the app fit together
- Package format: manifest v1.1 and the entrypoint format for each type
- Compatibility: where each type installs, per platform
- Submission guide Β· Contributor guide Β· Review policy
- Security model Β· Release process Β· Development & testing
- Quality report: heuristic scores for every resource
Catalog content is treated as untrusted data. The app never executes anything from a package. Packages that declare scripts are refused, and install paths are validated against traversal and symlinks. Every file is checked against the catalog's SHA-256 before and after writing, and any resource whose hash doesn't match is quarantined. Please report vulnerabilities privately through GitHub Security Advisories. See SECURITY.md.
Skilldeck was designed and built with Claude using Claude Code: the macOS app, the validation and release tooling, and the launch catalog of 250 resources. Every resource was written for real day-to-day use, then held to the same validator and review bar as community submissions.
MIT Β© Seyhun AkyΓΌrek and contributors. Each resource declares its own license in manifest.yaml. All launch resources are MIT.


