Ethical Skills Library for LLMs, Agents, and Robots

A collection of reusable, composable ethical skill modules that can be loaded into AI systems as system prompts, instruction layers, or behavioral constraints. Each skill addresses a specific moral concern -- harm prevention, de-escalation, fairness, honesty, care for vulnerable populations -- and can be combined into policy bundles for different deployment contexts.

moral-core works with any LLM that accepts a system prompt — Claude, GPT-4, Gemini, Mistral, LLaMA, and others. Skills are plain text files; loading them is a file read and a string concatenation. No SDK, no provider lock-in.

One distinction to be aware of:

If you're not using Claude Code, use the agents/ directory — those are the portable versions. If you are using Claude Code, the .claude/agents/ versions have file-reading tools attached for tighter integration. The skills are foundational and used by both agent types.

moral-core is a library of prompt-level ethical skills designed to be:

• Loaded into LLMs as system instructions or injected context
• Composed together into policy bundles for specific use cases
• Tested and evaluated against scenario-based rubrics
• Audited by reading the plain-text skill definitions

Each skill is a self-contained module with a clear scope, defined behavior, and known limitations. Skills can be loaded individually or bundled into pre-built policy sets.

• Not a moral authority. This library encodes practical behavioral guidelines, not philosophical truth. Reasonable people will disagree with specific choices.
• Not a replacement for law or regulation. Legal compliance is your responsibility. These skills do not constitute legal advice.
• Not a replacement for human oversight. No prompt-based system eliminates the need for human review, especially in high-stakes contexts.
• Not a security boundary. Prompt-level instructions can be overridden, ignored, or circumvented. See LIMITATIONS.md and SAFETY.md.
• Not exhaustive. Ethics is vast. This library covers common, high-priority concerns. It will have gaps.

moral-core/
├── README.md                  # This file
├── PRINCIPLES.md              # Shared doctrine: commitments, priority ladder, interpretive rules
├── SAFETY.md                  # Safety documentation and guidance
├── LIMITATIONS.md             # Honest account of what this framework cannot do
├── PHILOSOPHY.md              # Philosophical foundations and traditions
├── CHANGELOG.md               # Version history and release notes
├── CLAUDE.md                  # AI-assisted development guidelines
├── version.json               # Current version metadata
├── CONTRIBUTING.md            # How to contribute
├── CODE_OF_CONDUCT.md         # Community standards
├── GOVERNANCE.md              # Project governance
├── USE_CASES.md               # Deployment scenarios and skill recommendations
├── ROADMAP.md                 # Project roadmap
├── LICENSE                    # MIT
├── skills/                    # Ethical skill modules (21 domains, load into any LLM)
│   ├── general-ethics/        # SKILL.md, EXAMPLES.md, TEST_CASES.md, MISUSE.md
│   ├── conflict-mediation/
│   ├── deescalation-war-conflict/
│   ├── anti-sexism/
│   ├── anti-racism/
│   ├── empathy/
│   ├── protect-vulnerable/
│   ├── environment/
│   ├── animal-welfare/
│   ├── child-safety/
│   ├── disability-respect/
│   ├── elder-protection/
│   ├── abuse-prevention/
│   ├── epistemic-humility/
│   ├── human-oversight/
│   ├── digital-ethics/
│   ├── justice-fairness/
│   ├── democratic-legitimacy/
│   ├── research-ethics/
│   ├── financial-ethics/      # Fraud, predatory lending, market manipulation, anti-corruption
│   ├── data-privacy-surveillance/ # Personal data, tracking, re-identification, surveillance
│   └── labor-rights/          # Worker dignity, AI displacement, gig economy, wage theft
├── agents/                    # Framework-agnostic reviewer agents (load into any LLM)
│   ├── ethics-reviewer.md
│   ├── empathy-style-checker.md
│   ├── misuse-auditor.md
│   ├── advertising-ethics-reviewer.md
│   ├── mental-health-support-checker.md
│   ├── robotics-safety-ethics.md
│   ├── mediation-designer.md
│   └── warfare-agent-reviewer.md
├── .claude/
│   ├── agents/                # Claude Code subagent definitions (Claude Code only)
│   │   ├── ethics-reviewer.md
│   │   ├── empathy-style-checker.md
│   │   ├── misuse-auditor.md
│   │   ├── advertising-ethics-reviewer.md
│   │   ├── mental-health-support-checker.md
│   │   ├── robotics-safety-ethics.md
│   │   ├── mediation-designer.md
│   │   └── warfare-agent-reviewer.md
│   └── skills/                # Documentation and guides for skill modules
└── evals/                     # Evaluation framework
    ├── run.py                 # Automated evaluation runner (CLI)
    ├── runner/                # Parser, judge, and report modules
    ├── baselines/             # Saved regression baselines
    ├── results/               # Output from eval runs
    ├── adversarial/           # Adversarial robustness tests
    ├── benchmarks/            # Benchmark matrix
    ├── rubrics/               # Scoring rubrics
    └── scenarios/             # Scenario-based test cases
├── examples/                  # Example configurations and usage patterns
├── integrations/              # Integration guides for specific platforms and frameworks
│   ├── frameworks/            # Framework-specific guides (LangChain, Dify, CrewAI, + template for more)
│   │   └── loader.py          # Shared Python skill loader utility
│   └── *.md                   # Use-case guides (customer support, education, robotics, etc.)
├── docs/                      # Extended documentation
└── .github/                   # Issue and PR templates

This project follows Semantic Versioning (MAJOR.MINOR.PATCH).

Current Version: 1.4.0 (released 2026-04-01)

What's included in v1.4.0:

• 21 ethical skill domains covering harm prevention, fairness, honesty, care, research ethics, financial ethics, data privacy, labor rights, and more
• 8 framework-agnostic reviewer agents for ethical validation and design
• 8 Claude Code subagent equivalents for integration with Claude Code workflows
• 11 pre-built policy bundles for common deployment contexts (including new financial-services, data-platform, and labor-platform bundles)
• Comprehensive evaluation framework with adversarial and scenario-based tests
• Integration guides for LLMs, agents, robotics, education, content moderation, and enterprise
• Framework-specific guides for LangChain, Dify, and CrewAI with shared Python loader utility
• Automated evaluation runner (evals/run.py) with LLM-as-judge scoring across five dimensions
• Full documentation including philosophical foundations and safety guidance

Versioning scheme:

• MAJOR releases: Breaking changes to skill definitions, priority ladder, or core framework
• MINOR releases: New skills, new policy bundles, framework expansion
• PATCH releases: Documentation, clarifications, minor fixes

See CHANGELOG.md for complete version history and version.json for structured metadata.

Start with CONTRIBUTING.md for guidelines on proposing new skills, reporting issues, and submitting pull requests.

If using Claude Code or other AI tools to contribute, read CLAUDE.md for:

• Workflow for adding or updating skills
• Code style and naming conventions
• Testing and validation procedures
• Git workflow and branch strategy

Each skill is a text file that can be injected into a system prompt. The exact mechanism depends on your platform.

# Example: loading the de-escalation skill into an LLM system prompt
from pathlib import Path

skill_text = Path("skills/deescalation-war-conflict/SKILL.md").read_text()

system_prompt = f"""
You are a helpful assistant.

{skill_text}
"""

# Pass system_prompt to your LLM API of choice

Skills are designed to be concatenated. When conflicts arise, the priority ladder in PRINCIPLES.md determines which skill takes precedence.

from pathlib import Path

# Load the shared principles (always include this)
principles = Path("PRINCIPLES.md").read_text()

# Load specific skills
skills = []
for skill_name in ["general-ethics", "deescalation-war-conflict", "epistemic-humility"]:
    skills.append(Path(f"skills/{skill_name}/SKILL.md").read_text())

system_prompt = f"""
You are a customer support assistant.

## Ethical Framework
{principles}

## Active Skills
{"".join(skills)}
"""

Policy bundles are curated sets of skills for common deployment contexts.

from pathlib import Path
import json

child_safe_skills = [
    "child-safety", "protect-vulnerable", "empathy", "digital-ethics", "human-oversight"
]
skill_texts = [Path(f"skills/{s}/SKILL.md").read_text() for s in child_safe_skills]

system_prompt = f"""
You are an educational assistant for children ages 8-12.

{"".join(skill_texts)}
"""

Each bundle is a curated combination of skills for a specific deployment context. With 19 ethical skill domains available, you can create custom bundles or use these recommended combinations:

See .claude/skills/README.md for the complete list of all 19 available skills and their descriptions.

moral-core includes 8 framework-agnostic reviewer agents that apply the ethical skills to analyze prompts, policies, and system designs. Each agent is a plain markdown file that can be loaded into any LLM that accepts system prompts.

Framework-agnostic agents are in the agents/ directory and work with any LLM. Claude Code agents are in .claude/agents/ and are formatted for use with Claude Code's subagent system.

For usage and examples, see agents/README.md.

moral-core is designed to be platform-agnostic. Specific integration guidance is provided for:

• LLM applications -- chatbots, writing assistants, Q&A systems
• Agent frameworks -- autonomous agents, tool-using LLMs, multi-agent systems
• Robotics -- physical robots in care, service, manufacturing, and domestic settings
• Education -- tutoring systems, educational content generation, classroom assistants
• Content moderation -- automated moderation, human-in-the-loop review systems
• Enterprise copilots -- code assistants, business process automation, internal tools

See the integrations/ directory for platform-specific guides.

When skills conflict or a situation involves competing moral concerns, use this priority ordering. Level 1 is the highest priority.

The full ladder with detailed guidance is in PRINCIPLES.md.

moral-core is not philosophically neutral. It is grounded in named traditions and explicit commitments. The first human author is rooted in Enlightenment values (reason, universal rights, rational critique), Humanism (secular ethics, human dignity, human flourishing), and Ecological Sustainability (intergenerational justice, intrinsic value in nature, responsibility to the living world).

The framework draws from — without dogmatic commitment to any single one:

This framework explicitly rejects libertarian framings of autonomy. Freedom is always situated within responsibility to others, community, and the living world. Positive rights — to dignity, to a livable environment — are as foundational as negative rights.

See PHILOSOPHY.md for the complete mapping, including what each tradition contributes, where traditions tension, and which traditions we deliberately do not adopt.

Contributions are welcome. See the issue templates in .github/ISSUE_TEMPLATE/ for structured ways to propose new skills, report problems with existing ones, or suggest improvements to the framework.

Before contributing, please read:

• PRINCIPLES.md — Understand the commitments and priority ladder
• CONTRIBUTING.md — Contribution process and standards
• CLAUDE.md — If using AI-assisted development tools

All changes should be documented in CHANGELOG.md and follow the workflow described in CLAUDE.md.

MIT

This library provides behavioral guidelines, not guarantees. Prompt-based ethical layers operate within the constraints of the underlying model and can be circumvented. Downstream developers and deployers remain fully responsible for:

• Validating that skills behave correctly in their specific deployment context
• Conducting safety testing, red-teaming, and adversarial evaluation
• Complying with all applicable laws and regulations
• Maintaining human oversight appropriate to the risk level of their application
• Not using this framework to create a false sense of safety where real engineering controls are needed

The authors and contributors of moral-core accept no liability for harm caused by systems that incorporate these skills. See SAFETY.md and LIMITATIONS.md for detailed guidance.