ellmos skills

Six-language documentation · Machine-readable context · 🗺️ Browse the skill library online — read and copy every public skill in the browser

Portable AI skill library for Claude Code-style SKILL.md workflows, Codex-compatible agent setups, BACH, AGY/Gemini, and other local-first LLM agent runtimes.

CI: Tests License: MIT Pytest: 145 passed Python: >=3.10 | 3.13 Privacy: Zero-Egress Security: Local-First Organization: ellmos-ai Umbrella: open-bricks Public Skills: 135 Catalog Tracked: 380 Skills LLM-Ready: llms.txt

[!NOTE] AI Agent & LLM Integration: This repository provides standardized SKILL.md files with YAML frontmatter that can be consumed directly by Claude Code, Codex, AGY/Gemini, and custom agent runtimes. See llms.txt for machine-readable context.

[!IMPORTANT] Reading a copy? The canonical, always-current version of this library lives at github.com/ellmos-ai/skills. Forks and mirrors are not updated automatically and may be many commits behind — check the source before relying on anything you read here.

Quick links: Start Here · Featured Skills · Skills · Skills Map · Security · Conventions · Changelog

This repository is the reusable skill catalog of the ellmos ecosystem. It contains standalone process skills, development workflows, research helpers, therapy-oriented methods, infrastructure playbooks, and utility tools in an Anthropic-compatible SKILL.md format. Each skill carries its own metadata directly in YAML frontmatter, so runtimes can inspect provenance, compatibility, and dependencies without a central registry.

System Architecture

flowchart TD
    Registry["Public Skill Registry (135 Catalog / 380 Tracked)"] --> Engine["ellmos Skill Runtime & Dispatcher"]
    
    subgraph Catalog ["11 Public Domains"]
        Assist["assist (20)"]
        Dev["dev (25)"]
        Edu["education (5)"]
        Game["game-dev (5)"]
        Infra["infrastructure (29)"]
        Prod["production (1)"]
        Res["research (1)"]
        Therapy["therapy (20)"]
        ThirdParty["third-party (2)"]
        Utils["utilities (26)"]
        Web["web (1)"]
    end
    
    Engine --> Catalog
    Catalog --> Artifacts["SKILL.md Specifications\n(YAML Frontmatter + Playbooks + Scripts)"]
    
    subgraph MultiAgentRuntimes ["Multi-Agent Execution Fabric"]
        ClaudeCode["Claude Code (~/.claude/skills)"]
        Codex["Codex (~/.codex/skills)"]
        AGY["Antigravity / Gemini"]
        BACH["BACH Text-OS"]
        LocalOllama["Ollama / Local LLM"]
    end
    
    Artifacts --> MultiAgentRuntimes
    
    subgraph QualityGates ["Quality & Integrity Gates"]
        STests["S-Tests (Static Validation)"]
        LTests["L-Tests (LLM Self-Experience)"]
        UTests["U-Tests (User Experience)"]
        PytestSuite["Pytest Suite (145 Passed / 186 Subtests)"]
    end
    
    Artifacts -.-> QualityGates

Multi-Agent Skill Discovery & Execution Lifecycle

sequenceDiagram
    autonumber
    actor Operator as Agent / Developer
    participant Dispatcher as ellmos Skill Engine
    participant Registry as Public Catalog (components.json)
    participant Gate as Privacy & Integrity Gate
    participant Runtime as Target Agent (Claude / Codex / AGY / BACH)

    Operator->>Dispatcher: Query skill by domain or task ID (e.g. dev/pipeline-optimizer)
    Dispatcher->>Registry: Lookup metadata, schema version & dependencies
    Registry-->>Dispatcher: Return category, frontmatter spec & file path
    Dispatcher->>Gate: Execute privacy & non-elevation boundary check
    Gate-->>Dispatcher: Boundary verified (Zero-Egress & User-Mode clean)
    Dispatcher->>Runtime: Mount SKILL.md & contextual playbooks into agent workspace
    Runtime->>Runtime: Execute structured workflow steps deterministically
    Runtime-->>Operator: Deliver artifact, verification log & status receipt

Start Here

Need File or command
Browse all public skills skills/
See a tree map of every tracked skill SKILLS-MAP.md
Understand the SKILL.md schema docs/CONVENTIONS.md
Machine-readable catalog index registry/components.json
Security policy & boundary guarantees SECURITY.md
Browse by category skills/ (one subfolder per category)
Use a skill Copy skills/<category>/<name>/ into your agent's skills directory (e.g. ~/.claude/skills/)
Review public changes CHANGELOG.md
Give crawlers and LLM agents a compact map llms.txt

Catalog Snapshot

The current public catalog contains 135 public runtime skills (380 tracked across local suites):

Category Count Focus
assist 20 User-neutral methods for office work, notes, household planning, contacts, health-information organization, media and inventory exports, voice workflows, travel, weather, calendars, and transcription
dev 25 Development protocols, debugging, bug sweeps, pipeline renovation, migration, documentation, plugin systems, and repository publication
education 5 Academic planning, source-based learning, exam preparation, worksheet generation, and user-neutral teaching and support planning
game-dev 5 Blender, Roblox, Rojo, Studio, asset safety, and game-design workflows
infrastructure 31 Portable AI setup, system onboarding, skill landscape management, automation self-care, semantic persona routing, provider-neutral config sync and agent boot bridges
production 1 Text production router: general texts, narrative stories, PR with a local LaTeX press-release compiler
research 1 Research-agent workflow support
therapy 20 German-language psychoeducation and counseling method playbooks
third-party 3 Curated externally authored skills redistributed under verified licenses
utilities 26 Batch operations, thinking frameworks, decision briefings, document chunking, encoding repair, video transcripts, private-mail drafting, job-application support, user-model tooling, and German-law and German-tax first-look pointer skills
web 1 Web-reading protocol support

Featured Skills

Some skills are especially useful as entry points because they coordinate other tools, prevent messy agent workflows, or turn local procedures into repeatable playbooks:

Skill Why it stands out
skill-explorer Meta-skill for managing the skill landscape: audits existing skills, clusters them into families, researches external skills/plugins, and installs only after safety review and explicit approval.
model-strategy Multi-model routing for Claude, Codex, Gemini, and Ollama with score-based selection, delegation paths, escalation triggers, and cost/quality tradeoffs.
pipeline-optimizer Six-step renovation protocol for existing project folders, documentation systems, and software stacks; designed to avoid duplicate standards and broken workflows.
github-repo-care Publication and maintenance gate for GitHub repos: local rules, locks, .gitignore, privacy checks, README/i18n, releases, and repository metadata.
mcp-config-sync Provider-neutral MCP discovery and sync planning across user-selected providers and app classes; no implicit hub.
video-transcriber Extracts video subtitles/transcripts plus metadata (supports YouTube sources) into Markdown, JSON, or plain text so video analysis starts from source-backed text.
rbx-studio Roblox Studio operation: Explorer/Play-Test workflow, Rojo scene-vs-code hookup, AI control of Studio via MCP, and a mandatory malware scan for Creator Store assets.
decision-briefing Turns many open decisions into a numbered A/B/C/D briefing with recommendations, accepts batch replies, and records the chosen outcomes.
bugsweep Systematic bug-sweep protocol with a codebase-scaled target count, doubling escalation, area tracking, and completion verification -- turns ad hoc bug hunting into a repeatable, measurable pass.
plugin-system Generic auto-discovery plugin system for Python applications: zero dependencies (stdlib only), validation, and fault tolerance for turning your own scripts into a pluggable architecture.
bilingual-doc-sync Keeps parallel-language documents (papers, READMEs, SKILL.md/SKILL.en.md pairs) in sync: detects missing translations and section drift, plus an expansion audit for whether a document deserves further languages.
trampelpfadanalyse Empirical baseline-intervention-retest method for checking whether an agent convention or README rule is actually visible and followed, using isolated sandbox subagents to measure whether a doc change changed behavior.
law-checker Pointer skill to the standalone ellmos-ai/law-checker module: source-grounded AI first-look legal assessments for German law with a statute registry and a statute-embodiment agent -- AI orientation only, not a substitute for a lawyer.
steuer-assistent Pointer skill to the standalone ellmos-ai/steuer-assistent module: an offline-first local worksheet for German employee income-related expenses (Werbungskosten) -- not tax advice, no tax-return filing.
worksheet-generator Pointer skill to the standalone ellmos-ai/worksheet-generator module: generates individualized worksheets from a support goal, level, and age for educational/therapeutic professionals, with bring-your-own ICF references -- a material generator, not a therapy program.
research-agent Self-contained scientific literature workflow around PubMed and arXiv (pure Python stdlib) -- turns ad hoc paper hunting into a repeatable, source-backed research pass, fully portable without the ellmos ecosystem.
agent-config-sync Discovers provider/app-class surfaces and plans user-selected MCP, skill and rule-file truth topologies.
agents-bridge Portable provider-neutral file bridge: captures explicit boot/truth graphs, separate memory silos, messaging, presence and locks, then previews, restores, verifies or rolls back a privacy-gated instance package.
automation-self-care Builds a provider-neutral maintenance core set for scheduled LLM tasks and desktop-app automations with native readback, rollback and cross-system coverage.
semantic-persona-routing Routes requests through coordinator roles, experts and verified live skill endpoints while keeping persona overlays separate from capabilities and permissions.
build-your-users-mind Public, user-neutral pointer for building an authorized empirical preference model; personal profiles and evidence remain private.
dev-soft-agent Standalone development-automation pipeline (code analysis, task engine, policies, prompt templates) in zero-dependency Python -- a complete dev-agent workflow without external services.
llm-text-hygiene Removes AI traces and chat residue from finished texts and handles AI-disclosure levels -- keeps published documents clean of LLM artifacts.
idea-mining Distinctive multi-technique method for mining ideas out of stuck problems -- the structured alternative to free-form brainstorming when a project has hit a wall.
skill-extractor Extracts a reusable skill from a chat history (current session or transcript files) -- turns what just worked into a portable playbook, or improves an existing skill from the evidence.
workflow-extract Builds an automation from a chat history or from existing automation prompts of another agent system -- conversations become repeatable workflows.
ai-portable-setup Creates a portable offline AI working environment on a USB stick or any drive: local LLM models and a RAG pipeline, no cloud required.
bewerbungsexperte End-to-end job-application support: job-ad analysis, CV/LinkedIn optimization, cover letters, plus a database/folder-driven ASCII CV generator (German-language focus).
therapy/ collection The 19-skill therapy family (flagships: cognitive-restructuring, motivational-interviewing) -- evidence-cited, bilingual, ethics-gated psychoeducation and counseling method playbooks; the deepest coherent block of the library.

Public/Private Boundary

Public skill folders contain only portable methods and neutral assets. App- or host-specific adapters, accounts, databases, local paths, real user data, and personal defaults belong in a separate private profile or private fork. The repository privacy gate rejects concrete user-home paths, known private hosts, token patterns, and accidentally tracked ignored files.

foerderplaner follows this model explicitly: the public skill covers teaching and support planning only. General report generation is a separate public project, report-forge. Personal support-report templates and profiles are not part of this repository.

The same boundary applies elsewhere: build-your-users-mind and decision-avatar are the public user-model cores; named personal avatars remain private. Store-wave operator workflows are private-only and are not shipped. law-checker is the public legal orientation module; private legal-department workflows are not shipped.

The public catalog contains only Ellmos-authored skills. Third-party skills are not republished under an Ellmos author name. The public registry/components.json is therefore a minimal discovery index; internal ownership assessments, privacy classifications and the full maintainer registry remain in a separate No-Push repository. Its non-circular source authority is the versioned registry/public-skill-files.json: Git generates and verifies it in checkouts, while gitless archives and enriched Plan-D projections use only its listed files and ignore local private extras.

Education Skills

Five institution- and user-neutral education skills. The public foerderplaner plans teaching and support measures; it does not generate personal support reports.

Skill What it does
academic-study-control Semester planning, deadline tracking, exam registration, re-enrollment, mail/portal checks, and calendar reminders with source verification and privacy guardrails.
academic-study-learn Five-phase source-based learning cycle: clarify objective → extract key ideas → build glossary → transfer/apply → retrieval practice with gap tracking.
academic-study-test Five test modes (quick test, exam block, oral exam, assignment training, error diagnosis) with a rubric-based assessment system and a strict ethics boundary against live-exam support.
foerderplaner User-neutral teaching and support planning with goals, measures, differentiation, observation criteria, and review points; no report generator.
worksheet-generator Differentiated worksheets and learning materials based on a learning goal and level.

Repository Structure

skills/
  <category>/
    <skill-name>/
      SKILL.md              # Definition, frontmatter, usage workflow
      scripts/              # Optional executable helpers
      references/           # Optional supporting documents
  _templates/               # Templates for new skills
docs/
  CONVENTIONS.md            # Frontmatter specification
registry/components.json    # Minimal public catalog index
registry/public-skill-files.json # Public source authority for gitless copies
llms.txt                    # Compact project map for LLM crawlers

Skill Metadata

Every SKILL.md declares whether it works standalone, whether it is compatible with BACH, and where it came from:

standalone: true
bach_compatible: true
bach_origin: true
provenance:
  origin: "bach"
  origin_path: "system/skills/therapie/"
  origin_version: "1.0.0"
  last_sync_from_origin: "2026-03-12"
  last_sync_to_origin: null
  local_changes_since_sync: false

Supported skill types are skill, agent, expert, service, protocol, and tool.

Validation

Pull requests and pushes that change a public SKILL.md run the complete static S-test gate. Run the same tracked-skill check locally with:

python testing/skill_tester.py batch --type static --ci

With pre-commit installed, enable the repository hook once with pre-commit install. It applies the same gate only to changed SKILL.md files before a commit.

External Evaluations

Independent, third-party A/B evaluations of individual skills, referenced here as they become available (not run or commissioned by this project):

Search Context

Use this repository when searching for:

  • ellmos skills
  • ellmos-ai/skills
  • agent skill library
  • SKILL.md catalog
  • portable AI skills
  • Claude Code SKILL.md library
  • Codex skills library
  • Claude Code and Codex skills
  • local-first LLM agent skills
  • BACH skill catalog
  • Anthropic-compatible skills

The name is intentionally generic, so use the canonical repository string ellmos-ai/skills when linking or indexing this project. It is a reusable skill catalog, not an MCP server, hosted SaaS marketplace, prompt pack, or private skill installer.

Ecosystem & Sibling Projects

Project Organization Role
BACH ellmos-ai Full text-based LLM operating system
ellmos-core ellmos-ai Core runtime primitives and execution fabric
ellmos-controlcenter-mcp ellmos-ai Unified tool and profile gateway MCP server
system-explorer ellmos-ai Agent fleet composition and system exploration
workflowhooker ellmos-ai Transactional workflow hook dispatcher
sqlite-transit-sync ellmos-ai Offline-first transit sync and snapshot retention
DevCenter dev-bricks Desktop developer workstation suite
CodeBox dev-bricks Multi-language code editor and sandbox environment

License

MIT License. See LICENSE.

Liability

This project is an unpaid open-source donation. Liability is limited to intent and gross negligence under Section 521 of the German Civil Code. Use at your own risk. No warranty, maintenance guarantee, availability guarantee, or fitness-for-purpose guarantee is provided.