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.

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 Navigation

• 01. System Architecture & Quick Reference
• 02. Multi-Agent Skill Discovery & Execution Lifecycle
• 03. Target Personas & Discoverability
• 04. Comparative Matrix vs Alternatives
• 05. ASCII Four-View Architectural Topology
• 06. Start Here & Quickstart Playbook
• 07. Catalog Snapshot & Domain Distribution
• 08. Featured Skills & Real-World Workflows
• 09. Public/Private Boundary & Isolation
• 10. Education Skills
• 11. Repository Structure & Directory Layout
• 12. Skill Metadata & Frontmatter Anatomy
• 13. Validation & Automated Quality Gates
• 14. Search Context & High-Intent Indexing
• 15. Governance & Runtime Invariants Matrix
• 16. Ecosystem & Sibling Projects
• 17. Third-Party Licenses, Level 1 SBOM & RunAsInvoker
• 18. Security Policy, § 521 BGB Statutory Disclaimer & 48h SLA

---

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.

1. System Architecture & Quick Reference

flowchart TD
    Registry["Public Skill Registry (142 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 (32)"]
        Prod["production (1)"]
        Res["research (1)"]
        Therapy["therapy (20)"]
        ThirdParty["third-party (3)"]
        Utils["utilities (29)"]
        Web["web (1)"]
    end
    
    Engine --> Catalog
    Catalog --> Artifacts["SKILL.md Specifications<br/>(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 (455 Passed / 186 Subtests)"]
    end
    
    Artifacts -.-> QualityGates

Loading

2. Multi-Agent Skill Discovery & Execution Lifecycle

sequenceDiagram
    autonumber
    actor Operator as "Agent / Developer"
    participant Dispatcher as "ellmos Skill Engine"
    participant Registry as "Public Catalog"
    participant Gate as "Privacy and Integrity Gate"
    participant Runtime as "Target Agent (Claude, Codex, AGY, BACH)"

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

Loading

3. Target Personas & Discoverability

ellmos-skills serves four core technical personas across the autonomous agent and platform engineering spectrum:

[PERSONA-01] Autonomous AI Agents & Swarms

• Profile: Multi-agent frameworks, orchestrators, and autonomous coding loops (Claude Code, AGY, Codex, BACH).
• Core Pain Point: Hallucinated workflow steps, fragmented tools, non-deterministic task execution across agent boundaries, and state drift.
• How ellmos-skills Solves It: Standardized SKILL.md format with strict YAML frontmatter, self-contained playbooks, deterministic input/output contracts, and cold-start portability.
• Key Workflows & Skills: infrastructure/agents-bridge, dev/pipeline-optimizer, utilities/generalizer, infrastructure/work-autonomous.

[PERSONA-02] Enterprise DevOps & Platform Engineers

• Profile: Infrastructure architects and platform teams managing developer fleets, scheduled automations, and CI pipelines.
• Core Pain Point: Re-inventing operational maintenance, brittle ad-hoc scripts, lack of cross-host parity, and chaotic repository housekeeping.
• How ellmos-skills Solves It: Ready-to-mount, battle-tested engineering playbooks for refactoring, Git hygiene, and pipeline optimization without vendor tool sprawl.
• Key Workflows & Skills: dev/project-bootstrapper, dev/pipeline-bootstrapper, utilities/folder-organization, dev/github-repo-care.

[PERSONA-03] Local-First, Privacy & SecOps Specialists

• Profile: Security officers, compliance auditors, and privacy engineers in regulated or air-gapped environments.
• Core Pain Point: Accidental cloud telemetry leaks, unverified transitive packages, supply-chain vulnerabilities, and elevated execution privileges in agent tools.
• How ellmos-skills Solves It: Strict Zero-Egress (INV-LOCAL-01), automated static privacy boundary gates (testing/privacy_gate.py), user-mode execution (RunAsInvoker), 0% copyleft, and Level 1 SBOM transparency.
• Key Workflows & Skills: infrastructure/privacy-gate, utilities/secret-redactor, SECURITY.md, THIRD_PARTY_LICENSES.md.

[PERSONA-04] Domain Skill Authors & Research Engineers

• Profile: AI researchers, methodologists, and domain specialists creating reproducible agent capabilities.
• Core Pain Point: Lack of conventions, inconsistent multi-language handling, and missing testing harnesses for custom skills.
• How ellmos-skills Solves It: Formal schema specification (docs/CONVENTIONS.md), public registry generator (registry/components.json), and three-tier S/L/U testing framework.
• Key Workflows & Skills: schemas/assist-v1.schema.json, docs/CONVENTIONS.md, testing/skill_tester.py, research/research-agent.

Target Personas Summary Matrix

Persona	Core Pain Point	How ellmos-skills Solves It	Key Workflows & Skills
Autonomous AI Agents & Swarms (Claude Code, AGY, Codex, BACH)	Hallucinated workflow steps, fragmented tools, and non-deterministic task execution across agent boundaries.	Standardized SKILL.md format with strict YAML frontmatter, self-contained playbooks, and deterministic input/output contracts.	infrastructure/agents-bridge, dev/pipeline-optimizer, utilities/generalizer
Enterprise DevOps & Platform Engineers	Re-inventing operational maintenance, brittle CI scripts, and chaotic repository housekeeping.	Ready-to-mount, battle-tested engineering playbooks for refactoring, Git hygiene, and pipeline optimization without tool sprawl.	dev/project-bootstrapper, dev/pipeline-bootstrapper, utilities/folder-organization
Local-First, Privacy & SecOps Specialists	Accidental cloud data leaks, unverified transitive packages, and elevated execution privileges in agent tools.	Strict Zero-Egress (INV-LOCAL-01), automated privacy boundary gates, user-mode execution (RunAsInvoker), and 0% copyleft.	infrastructure/privacy-gate, utilities/secret-redactor, SECURITY.md
Domain Skill Authors & Research Engineers	Lack of conventions, inconsistent multi-language handling, and missing testing harnesses for custom skills.	Formal schema specification (docs/CONVENTIONS.md), public registry generator (registry/components.json), and S/L/U testing framework.	schemas/assist-v1.schema.json, docs/CONVENTIONS.md, testing/skill_tester.py

High-Intent Search Keyword Matrix

Category	Primary Search Terms (English)	Relevanz & Suchbegriffe (Deutsch)
Agent Frameworks	claude code skills, codex agent skills, anthropic skill md standard, agy gemini agent skills	Claude Code Skills, Codex Agenten Skills, Anthropic SKILL.md Standard, Gemini Agenten Skills
Skill Architecture	portable ai skills, reusable agent playbooks, agent skill library, deterministic llm workflows	Portable KI Fähigkeiten, Wiederverwendbare Agenten Playbooks, KI Skill Bibliothek, Deterministische Workflows
Privacy & Security	zero-egress ai skills, local-first agent library, fail-closed privacy gate, user-mode agent execution	Zero-Egress KI Skills, Lokale Agenten Bibliothek, Fail-Closed Datenschutz Gate, Unprivilegierte Ausführung
Ecosystem & Catalog	bach skill catalog, multi-agent skill library, offline ai tools, open source agent skills	BACH Skill Katalog, Multi-Agenten Skill Bibliothek, Offline KI Werkzeuge, Open Source Agenten Skills

Chinese users can also search and install skills through Skills宝.

4. Comparative Matrix vs Alternatives

ellmos-skills provides a vendor-neutral, portable skill standard optimized for local-first, multi-agent operations. The following matrix illustrates how it compares against common architectural approaches across ten critical dimensions:

Evaluation Dimension	ellmos-skills	Ad-Hoc System Prompts	Tool/Function Calling Only	Centralized Cloud Hubs	Heavyweight Frameworks (LangChain/CrewAI)
1. 100% Local-First & Zero-Egress	PASS (Zero telemetry)	PASS (Local text)	⚠️ Depends on API provider	❌ Cloud account & telemetry required	⚠️ Complex dependencies & telemetry risks
2. Standardized Format (SKILL.md)	PASS (Anthropic standard + YAML)	❌ Freeform unstructured text	❌ JSON schema only, no playbooks	⚠️ Vendor-specific manifests	❌ Python/JS code abstractions
3. Multi-Agent Portability	PASS (Claude, Codex, AGY, BACH)	⚠️ Prompts drift across models	⚠️ Custom wrappers per provider	❌ Locked to proprietary cloud	⚠️ Framework-specific lock-in
4. Step-by-Step Playbooks	PASS (Structured phases & gates)	❌ Inconsistent instruction adherence	❌ Single-step tool execution only	⚠️ Varies wildly by author	⚠️ Python DAG logic required
5. Privacy & Leakage Gates	PASS (Automated static gate)	❌ None	❌ None	⚠️ Cloud moderation filters only	❌ None built-in
6. Automated Test Suite	PASS (455 Pytest tests, 186 subtests, 100% green)	❌ Untested	⚠️ Unit tests for API calls only	❌ No public test harness	⚠️ Framework unit tests only
7. Multi-OS CI Parity	PASS (Linux, Windows, macOS)	N/A	⚠️ Provider-dependent	❌ Hosted service	⚠️ OS compatibility quirks
8. Machine-Readable Catalog	PASS (registry/components.json)	❌ Unindexed	❌ Dynamic API introspection only	⚠️ Proprietary API queries	❌ Code module imports
9. Open-Source & 0% Copyleft	PASS (100% MIT License)	N/A	N/A	❌ Proprietary platform terms	⚠️ Mix of licenses
10. Zero External Runtime Deps	PASS (Python stdlib only)	PASS	⚠️ HTTP/JSON libraries needed	❌ Cloud client SDKs required	❌ Massive dependency trees

5. ASCII Four-View Architectural Topology

+----------------------------------------------------------------------------------------------------+
| VIEW 1: RUNTIME DISPATCH & DISCOVERY PLANE                                                         |
|                                                                                                    |
|  [ Operator / LLM Agent ]                                                                          |
|            |                                                                                       |
|            v (Task Intent / Query)                                                                 |
|  +----------------------------------------------------------------------------------------------+  |
|  | ellmos Skill Runtime & Dispatcher (Pure Python Stdlib)                                        |  |
|  |  * Category Resolver     * Frontmatter Validator    * Dependency & Provenance Introspector   |  |
|  +----------------------------------------------------------------------------------------------+  |
|            |                                            |                                          |
|            v (Discovery Query)                          v (Catalog Index Readback)                 |
|  +-----------------------------------+        +-------------------------------------------------+  |
|  | Public Catalog Registry           |        | llms.txt & Context Index                        |  |
|  | (registry/components.json: 142)   |        | (Machine-readable markdown & category hierarchy)|  |
|  +-----------------------------------+        +-------------------------------------------------+  |
+----------------------------------------------------------------------------------------------------+
| VIEW 2: MULTI-AGENT CLIENT INTEGRATION LAYER                                                       |
|                                                                                                    |
|  [ Claude Code ]      [ OpenAI Codex ]      [ Antigravity / Gemini ]     [ BACH Text-OS / Local ]  |
|  (~/.claude/skills)   (~/.codex/skills)     (~/.gemini/.../skills)       (system/skills/...)       |
|         \                   |                        |                          /                  |
|          +------------------+------------------------+-------------------------+                   |
|                             | Uniform Materialization                                              |
|                             v                                                                      |
|  +----------------------------------------------------------------------------------------------+  |
|  | Standardized SKILL.md Execution Surface                                                      |  |
|  |  * Anthropic YAML Frontmatter Contract   * Step-by-Step Playbooks   * S/L/U Verification Gates  |  |
|  +----------------------------------------------------------------------------------------------+  |
+----------------------------------------------------------------------------------------------------+
| VIEW 3: GOVERNANCE & SECURITY BOUNDARY (FAIL-CLOSED)                                               |
|                                                                                                    |
|  +----------------------------------------------------------------------------------------------+  |
|  | Static Privacy Gate (testing/privacy_gate.py)                                                 |  |
|  |  * Zero-Egress Assertion (INV-LOCAL-01)          * Non-Elevation / RunAsInvoker (INV-UNPRIV-03) |  |
|  |  * Rejection of Concrete User Homes & Hostnames * Rejection of API Keys, Tokens & Private Data |  |
|  +----------------------------------------------------------------------------------------------+  |
|            |                                            |                                          |
|            v (Compliance Audit)                         v (Public/Private Boundary)                |
|  +-----------------------------------+        +-------------------------------------------------+  |
|  | Level 1 SBOM Invariant Matrix     |        | No-Push Boundary Separation                     |  |
|  | (THIRD_PARTY_LICENSES.md/.txt)    |        | (Public generic methods vs Private profiles)    |  |
|  +-----------------------------------+        +-------------------------------------------------+  |
+----------------------------------------------------------------------------------------------------+
| VIEW 4: DOMAIN TOPOLOGY & SKILL ASSET ANATOMY                                                      |
|                                                                                                    |
|  skills/                                                                                           |
|  +-- assist/ (20)        +-- dev/ (25)            +-- education/ (5)       +-- game-dev/ (5)       |
|  +-- infrastructure/ (32)+-- production/ (1)      +-- research/ (1)        +-- therapy/ (20)       |
|  +-- third-party/ (3)    +-- utilities/ (29)      +-- web/ (1)                                     |
|                                                                                                    |
|  [ Individual Skill Anatomy ]:                                                                     |
|  +-- SKILL.md (Frontmatter + Playbook)   +-- scripts/ (Helpers)   +-- references/ (Evidence/Docs)  |
+----------------------------------------------------------------------------------------------------+

6. Start Here & Quickstart Playbook

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
Level 1 SBOM Plain-Text Companion	THIRD_PARTY_LICENSES.txt
Formal copyright & attribution notice	NOTICE
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

7. Catalog Snapshot & Domain Distribution

The current public catalog contains 142 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	32	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	29	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

8. Featured Skills & Real-World Workflows

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.
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.
lebende-verfassung	Operationalized constitutional superposition ('Position of the Unborn'): gives future and unborn generations a formal algorithmic Rawlsian veto right against short-term present-day optimizations via a 5-CORE review architecture and mandatory counterfactual analysis.
work-autonomous	Proof-based non-termination protocol (WAAFAP) against agentic laziness: inverts the stopping condition so that concluding an autonomous loop requires falsifiable proof of task absence ('quit requires proof of inactivity').
piggyback-hosting	Zero-state privacy deployment pattern (Huckepack-Hosting): runs relational SQLite in the visitor's browser via SQLite-WASM/OPFS with client-side BYOK, eliminating server databases, user accounts, and GDPR liability by design.
software-in-worten	Bidirectional UI-prompt synthesis ('The click is the prompt'): typed ASCII blueprint format with a 4D element legend that bridges GUI design and agent instructions without requiring a frontend build step.
metacognitive-injectors	Neuropsychological executive control functions (Miyake inhibition, working memory buffer, mental rehearsal) integrated into agent runtime preflight checks to prevent sycophancy and premature termination.
paveman	Deterministic, model-free rule compression: shrinks large Markdown rule and memory files by up to 40% token volume without LLM inference, hallucinations, or semantic drift.
wayfinding-routing	Universal nautical navigation protocols (Celestial Fix, Dead Reckoning) for disoriented AI agents: provides recovery heuristics and state reconstruction when trapped in context drift or loops.
condition	Declarative condition-gate DSL for prompts: encapsulates preconditions, milestones, and ordering dependencies into fail-closed gates within standard Markdown.
letter-hooker	Preflight bootloader for hook-less CLI agents: injects governance rules, memory traversal, and self-healing context before turns without requiring native JSON event hooks.
pingpong	Session-scoped radio station over shared synced folders: separates asymmetric roles (ListenSync listener vs. WriteSync delta sender) for asynchronous multi-agent coordination without central servers.
choose-your-orchestrator	Session-contract negotiator for multi-agent work: bounds orchestration topology, concurrency, model slots, and escalation triggers before execution starts.
reissverschluss-merge	Zipper merge protocol for heavily divergent branches: compares section-by-section using a decision table, with intent reconstruction ('rebuild instead of merge') as the final escalation stage.
migrate-rename	Evolutionary file and module renaming with wrappers and MOVED stubs: prevents breaking changes across agent fleets while references update organically through usage.
projekt-pipeline-umbrella	Taxonomic 2x2 pipeline compass (Greenfield vs. Brownfield x Project vs. Pipeline): prevents LLM scaffolding bias by routing to the right onboarding, bootstrapper, or renovation skill.
tidy-up	Deterministic 3-role session-end hygiene loop (Tasksolver, Writer, Maintainer): solves pending trivial tasks without agenda creep, syncs documentation to measured truth, and archives strays reversibly.
human-loop-audit	Asynchronous human-in-the-loop pipelining: while the user tests item N live, the agent pre-launches item N+1 and delegates fix workers for item N-1, eliminating idle wait times.
folder-organization	Semantic filesystem cleanup using Cut-and-Clue: separates active from legacy files with machine-readable pointer clues at the source, preserving taxonomies and audit logs.
iterative-bundle-selection	Reduces large candidate lists step by step through optional topic pools, filter stages, and repeatedly reshuffled bundles -- for selecting, filtering, or mixing skills, tasks, ideas, or files in groups without permanently discarding the rest.

9. Public/Private Boundary & Fail-Closed Isolation

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.

Both registry/components.json and SKILLS-MAP.md are generated. After any change below skills/, run python build_public_registry.py and python build_skills_map.py, review the diff and commit it. Nothing regenerates silently: the CI step Check public catalog outputs and the pre-commit hooks public-registry-current / skills-map-current fail loudly on a stale catalog.

10. 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.

11. Repository Structure & Directory Layout

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
NOTICE                      # Formal copyright notice and attribution
THIRD_PARTY_LICENSES.txt    # Plain-text Level 1 SBOM companion
llms.txt                    # Compact project map for LLM crawlers

12. Skill Metadata & Frontmatter Anatomy

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.

13. Validation & Automated Quality Gates

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):

• cloud-communication-protocols -- decimal.ai, tested 2026-08-08 on Gemini-3.6-flash, 22 cases: pass rate 22.7% -> 95.5% (+73pp), -14% tokens, security 15/15 checks (3/3).

14. Search Context & High-Intent Indexing

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.

15. Governance & Runtime Invariants Matrix

Every skill, utility script, and metadata manifest within ellmos-skills is governed by ten foundational invariants:

Invariant ID	Name	Operational & Architectural Scope	Verification & Compliance Status
INV-LOCAL-01	100% Local-First & Zero Egress	Network & Privacy	PASS — Pure Markdown and local Python scripts; zero telemetry and zero outbound network calls.
INV-PRIVACY-02	Fail-Closed Privacy Boundary Gate	Data Protection	PASS — Automated privacy gate rejects concrete user homes, host-scoped names, and token patterns.
INV-UNPRIV-03	Non-Elevation & RunAsInvoker	Process Execution	PASS — All workflows execute strictly in unprivileged user space; root or admin elevation is prohibited.
INV-SCHEMA-04	Deterministic Frontmatter & Schema Integrity	Contract Compliance	PASS — YAML frontmatter strictly conforms to docs/CONVENTIONS.md and schema validation gates.
INV-ISOLATION-05	Portable & Self-Contained Skill Anatomy	Modularity	PASS — Skills package their own playbooks, references, and scripts without implicit host couplings.
INV-PORTABLE-06	Multi-Agent Runtime Portability	Interoperability	PASS — Compatible with Claude Code, Codex, AGY/Gemini, BACH, and local Ollama runtimes.
INV-DISCOVERY-07	Public/Private Boundary Isolation	Boundary Hygiene	PASS — Public catalog (registry/components.json) exposes only verified non-sensitive discovery fields.
INV-PLATFORM-08	Multi-OS Parity & Platform Independence	Portability	PASS — Linux, Windows, and macOS validated identically via multi-OS CI matrix on Python 3.10 to 3.13.
INV-SYNC-09	Multi-Host Cloud Sync & Lock Resilience	Concurrency	PASS — Hardened .gitignore and cooperative lock protocols prevent cloud sync races and data corruption.
INV-SLA-10	48h Security Response & Triage Commitment	Governance	PASS — Formal response SLA: 48h initial acknowledgment, 5-day triage commitment via SECURITY.md.

16. Ecosystem & Sibling Projects Matrix

Project	Organization	Role
BACH	ellmos-ai	Full text-based LLM operating system
ellmos-controlcenter-mcp	ellmos-ai	Unified tool and profile gateway MCP server
agent-ops-stack	ellmos-ai	Git-fetch manifest lists this library as a skill-pack source
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
MarbleRun	ellmos-ai	Local-first automation framework for autonomous LLM agent chains
gardener	ellmos-ai	Curated cross-source memory index for agentic systems
usmc	ellmos-ai	Local SQLite memory and cross-agent context sharing
DevCenter	dev-bricks	Desktop developer workstation suite
CodeBox	dev-bricks	Multi-language code editor and sandbox environment

17. Third-Party Licenses, Level 1 SBOM & RunAsInvoker

ellmos-skills is committed to absolute transparency, clean-room software boundaries, and supply chain security:

• Zero External Runtime Dependencies: All core catalog tools, schema generators, and privacy boundary verifiers run strictly on the Python Standard Library (>=3.10).
• Permissive Third-Party Skills: Curated external skills in skills/third-party/ (grill-me, grilling) are licensed under the MIT License from upstream mattpocock/skills.
• Formal Attribution & Notice: Legal notices and open-source attribution are recorded in NOTICE.
• Level 1 SBOM Companions: Complete third-party license inventory and invariant mapping are documented in THIRD_PARTY_LICENSES.md and plain-text companion THIRD_PARTY_LICENSES.txt.
• 0% Copyleft: No GPL, AGPL, or LGPL components are bundled or required at runtime.
• Unprivileged User Mode (RunAsInvoker): All scripts operate strictly in user-mode space without elevation.
• 10 Governance & Runtime Invariants: Every release is validated against strict invariants (INV-LOCAL-01 to INV-SLA-10).
• Full dependency details, license texts, and invariant definitions are documented in THIRD_PARTY_LICENSES.md.

18. Security Policy, § 521 BGB Statutory Disclaimer & 48h SLA

Security Policy & 48h SLA Commitment

We take security and privacy boundary enforcement seriously across all agent runtimes:

• Binding 48-Hour Response SLA: Receipt of vulnerability reports is acknowledged within 48 hours.
• 5-Day Triage SLA: Comprehensive security assessment and remediation roadmap provided within 5 business days.
• Reporting Channels: Contact security@ellmos.ai with CC to support@lukasgeiger.com or open a private GitHub Security Advisory. Do NOT file public issues for security vulnerabilities.
• Complete policy details: SECURITY.md.

Statutory Notice & Liability Limitation (§ 521 BGB)

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

License & Contribution

ellmos-skills is licensed under the MIT License. Full attribution and copyright notices are maintained in NOTICE. Detailed development guidelines, invariants, and submission protocols are documented in CONTRIBUTING.md.