What Is an OKF Skill? — A Guide to AI Agent Skills in Open Knowledge Format
OKF skills are reusable knowledge packages that teach AI agents specific capabilities — not just facts. Think of them as executable knowledge: instructions, patterns, and workflows packaged in the Open Knowledge Format.
What Makes an OKF Skill Different
OKF skills are a specialized subset of OKF bundles. While a general OKF bundle informs — providing reference documentation, API specs, or conceptual knowledge — an OKF skill instructs. A skill tells an AI agent how to do something, step by step, with executable patterns.
Skills Teach Capabilities, Not Just Inform
A knowledge bundle about Python would describe the language's syntax, standard library, and ecosystem. A Python OKF skill, by contrast, would contain instructions for how an agent should write, review, or debug Python code — including code style rules, error-handling patterns, testing requirements, and file organization conventions. The skill teaches the agent a capability, not a subject.
Skills Are Structured for Agent Execution
OKF skills use a consistent structure that AI agents recognize as executable instructions:
- Calling conventions — how an agent should invoke the skill (e.g., "When the user asks to distill a video, load the Lineage Skill")
- Step-by-step procedures — numbered actions the agent follows
- Behavioral rules — constraints and preferences (e.g., "Always validate frontmatter before writing")
- Output templates — the exact format the agent should produce
- Examples — worked-through demonstrations of the skill in action
Skills Can Be Combined and Composed
OKF skills are composable. An agent can load multiple skills simultaneously, each contributing a different capability. For example, an agent might load a documentation skill (how to write docs), a frontmatter validation skill (how to check YAML), and a knowledge-distillation skill (how to extract key points from source material) — all working together in one session.
How OKF Skills Work
Structure: Markdown + YAML Frontmatter
Like all OKF bundles, skills are directories of markdown files with YAML frontmatter. The key difference is that a skill's index.md acts as a manifest describing the skill's purpose, invocation, and behavioral expectations:
---
title: My Agent Skill
description: Teaches an agent to perform domain-specific tasks
tags: [skills, okf, agent-capability]
version: "1.0.0"
author: Your Name
okf_version: "0.1"
type: skill
---
# My Agent Skill
## Purpose
This skill teaches the agent how to [perform a specific task].
## When to Use
Load this skill when the user asks about [topic] or requests [action].
## Instructions
1. First, [step one with details]
2. Then, [step two]
3. Finally, [step three]
## Rules
- Always [constraint one]
- Never [constraint two]
- Use [format/pattern] for output
## Examples
### Example 1: [Scenario]
[Walk through the skill in action] How Agents Load and Use Skills
Different AI agents load OKF skills in different ways:
Claude Code
Claude Code discovers skills from the ~/.claude/skills/ directory. Any OKF bundle placed in this directory is auto-loaded into Claude's context. Agents reference skills by their name or description in prompts:
# Install a skill
git clone https://github.com/JuneYaooo/lineage-skill.git ~/.claude/skills/lineage-skill
# Use it — Claude automatically detects available skills
claude "Distill this course transcript using the Lineage Skill" OpenCode / Codex
OpenCode and Codex (by Nous Research) support OKF skills through the .opencode/ directory. Skills can be symlinked or copied into the agent's knowledge directory and are discovered at session start.
Generic Agent Integration
Any LLM-based agent can consume an OKF skill by reading its index.md and following its instructions:
# Load skill instructions into agent context
cat ~/skills/my-skill/index.md
cat ~/skills/my-skill/instructions/*.md Skills vs. General Knowledge Bundles
| Aspect | Knowledge Bundle | OKF Skill |
|---|---|---|
| Purpose | Inform — provide reference knowledge | Instruct — teach a capability |
| Content | Facts, docs, API references, concepts | Procedures, rules, templates, calling conventions |
| Agent usage | Read and understand a domain | Execute step-by-step instructions |
| Tags | Domain-specific tags (e.g., python, api) | skills, claude-skills, okf-skills |
| Composability | Linked via cross-references | Loaded simultaneously, combined at runtime |
| Output | No expected output format | Defines structured output templates |
Some bundles blur the line — they contain both reference knowledge and procedural instructions. These hybrid bundles are tagged with both domain tags and skill tags.
How AI Agents Load and Use Skills
The most popular agent platforms use different mechanisms to discover and load OKF skills:
Claude Code Skill System
Claude Code has a first-class skill system. Skills are OKF bundles placed in ~/.claude/skills/. Claude discovers them automatically and users invoke them conversationally:
# ~/.claude/skills/ directory structure
~/.claude/skills/
├── lineage-skill/
│ ├── index.md
│ ├── instructions/
│ └── templates/
├── okf-knowledge-skill/
│ ├── index.md
│ └── concepts/
└── okf-frontmatter/
├── index.md
└── scripts/ Manual Loading
For agents without a native skill system, you can load skills by including their content in the system prompt or context window:
# Load a skill into any agent's context
SKILL_DIR="$HOME/skills/my-skill"
echo "=== Loading Skill: $(head -1 $SKILL_DIR/index.md) ==="
cat $SKILL_DIR/**/*.md How to Create Your Own OKF Skill
Creating an OKF skill follows the same process as creating a knowledge bundle, with a few skill-specific additions:
Step 1: Define the Capability
Clearly define what you want the skill to teach. What task should the agent be able to perform after loading the skill? Write this as a one-sentence purpose statement.
Step 2: Create the Skill Directory
mkdir my-agent-skill
cd my-agent-skill Step 3: Write index.md with Skill Metadata
Include type: skill in your frontmatter and tag it with skills:
---
title: My Agent Skill
description: Teaches agents to perform [specific capability]
tags: [skills, okf, your-domain]
type: skill
version: "0.1.0"
author: Your Name
okf_version: "0.1"
---
# My Agent Skill
## Purpose
[One sentence describing the capability]
## When to Use
[When should the agent activate this skill?]
## Instructions
[Step-by-step process]
## Rules
[Behavioral constraints]
## Examples
[Worked-through demonstrations] Step 4: Add Supporting Concepts
Create subdirectories for templates, examples, and reference materials that the skill needs:
my-agent-skill/
├── index.md
├── instructions/
│ ├── setup.md
│ └── execution.md
├── templates/
│ └── output-template.md
└── examples/
└── basic-usage.md Step 5: Publish and Submit
Push to GitHub and submit to BundleDex with the skills tag so other users can discover your skill.
Best OKF Skills in the BundleDex Index
These are the top OKF skills currently indexed on BundleDex, ranked by GitHub stars:
Distill videos, PDFs, transcripts, and notes into source-backed teacher Agent Skills.
skill: maintain repo docs under the Open Knowledge Format + fast doc/schema lookup (find_docs.py). Grep-first, script-fallback.
Janet builds and maintains a portable LLM wiki in plain Markdown using the Open Knowledge Format (OKF). Run her directly, call her as a subagent, or add her skills to your coding agent.
Окружение для работы с OKF. SKILL для соблюдения формата и упрощения работы, rag для поиска
A curated collection of reusable skills, workflows, and knowledge modules for building LLM-powered okf-compatible wiki systems.
Template de governança de IA
Cross-agent Claude Code / Codex skill to author, convert, and validate Markdown in Open Knowledge Format (OKF).
Agent skill: structured, resumable domain-elicitation interviews that emit an OKF knowledge bundle (glossary, tri-temporal events, rules, context map)
Store, query, and generate knowledge in the Open Knowledge Format (OKF) - markdown bundles with YAML frontmatter, trust signals, CLI, and LLM ingestion.
The OKF toolkit for Claude Code — author, maintain, validate & visualize Open Knowledge Format bundles. Claude Code plugin + skills.sh.
Claude Code plugin: author, explore, validate, and visualize Google's Open Knowledge Format (OKF) v0.1 bundles.
Browse all 583 bundles on BundleDex, or filter by the skills tag to discover more.
FAQ
What is an OKF skill?
An OKF skill is a specialized Open Knowledge Format bundle that teaches an AI agent a specific capability — like how to distill knowledge, validate YAML frontmatter, or manage documentation. Unlike a general knowledge bundle that only informs, a skill provides step-by-step instructions, behavioral rules, and output templates that the agent follows during execution.
How do I use an OKF skill with Claude Code?
Clone the skill's GitHub repository into ~/.claude/skills/. Claude Code automatically discovers skills in this directory on startup. You can then invoke the skill by name in your prompts. For example: "Use the Lineage Skill to distill this video transcript into structured OKF knowledge." Or simply ask Claude to perform the task — it will recognize which skill to load.
How do I create my own OKF skill?
Create a directory with an index.md containing YAML frontmatter (include type: skill and tag skills). Write clear step-by-step instructions, behavioral rules, and examples. Add supporting subdirectories for templates and reference materials. Publish to GitHub and submit to BundleDex. The OKF Skills Toolkit and okf-author bundles can help automate creation.
Are OKF skills free?
Yes. All OKF skills indexed on BundleDex are open-source and free to use under permissive licenses (MIT, Apache-2.0, or similar). They are hosted on public GitHub repositories with no paid tiers or subscription requirements.
What's the difference between an OKF skill and a Claude Code skill?
An OKF skill is the format-agnostic concept — any OKF-format bundle tagged with skills. A Claude Code skill is an OKF skill placed in the ~/.claude/skills/ directory. Claude Code skills are a subset of the broader OKF skill ecosystem. The same skill can work with Claude Code, OpenCode, Codex, or any other agent that reads OKF bundles.
Can I use OKF skills with non-Claude agents?
Yes. OKF skills are agent-agnostic. Any LLM-based agent can consume an OKF skill by reading its markdown files. OpenCode, Codex, Cursor, and custom agent frameworks all support OKF skill loading through their own discovery mechanisms.