OmniSkill

A super skill generator that turns CSV and Markdown datasets into ready-to-use Agentic-RAG skills with a single command.

Overview

OmniSkill analyzes your dataset directory and generates:

• SKILL.md — Skill specification document that tells LLMs how to use the knowledge base
• search.py — Standalone Python script for BM25-based retrieval against the dataset
• datasets/ — Symlinked source data for the generated skill

It also provides a CLI for manually creating skill scaffolding and searching knowledge bases.

Installation

Using uv (Recommended)

uv add omniskill

Using pip

pip install omniskill

From Source

git clone https://github.com/longcipher/omni-skill.git
cd omni-skill
uv sync --all-groups

Quick Start

Generate a Skill from a Dataset

Point OmniSkill at any directory containing CSV and/or Markdown files:

omniskill generate examples/backend-api-master

This analyzes the dataset files and produces a complete skill under skills/<dataset-name>/:

skills/backend-api-master/
  SKILL.md      # Skill specification for LLMs
  search.py     # Standalone search script
  datasets/     # Symlink to source data

Use the Generated Skill

The generated search.py can be run directly:

python skills/backend-api-master/search.py "API design best practices"

Or use the CLI:

omniskill search "API design best practices" --skill-dir skills/backend-api-master

Custom Options

# Custom skill name and output directory
omniskill generate my-datasets/ --name my-skill --output out/my-skill/

# Verbose mode shows dataset analysis details
omniskill generate my-datasets/ --verbose

Architecture

graph TD
    subgraph "1. Dataset Input"
        CSV[CSV Files] --> GEN[omniskill generate]
        MD[Markdown Files] --> GEN
    end

    subgraph "2. Generator"
        GEN --> ANALYZE[Analyze Dataset]
        ANALYZE --> GEN_SCRIPT[Generate search.py]
        ANALYZE --> GEN_MD[Generate SKILL.md]
        ANALYZE --> LINK[Symlink datasets/]
    end

    subgraph "3. Generated Skill"
        GEN_SCRIPT --> SEARCH_PY[search.py]
        GEN_MD --> SKILL_MD[SKILL.md]
        LINK --> DATASETS[datasets/]
    end

    subgraph "4. Runtime"
        SEARCH_PY --> ENGINE[SearchEngine]
        ENGINE --> INDEXER[CSV / MD Indexer]
        ENGINE --> BM25[BM25 Searcher]
        BM25 --> ASM[PromptAssembler]
        ASM --> |XML / Markdown / llms.txt| OUTPUT[Formatted Context]
    end

    SKILL_MD -. "instructs LLM" .-> SEARCH_PY

CLI Reference

generate — Generate a Skill from Datasets

omniskill generate <dataset-dir> [options]

Arguments:

• dataset-dir — Path to a directory containing CSV and/or Markdown files

Options:

• --name, -n — Skill name (defaults to the dataset directory name)
• --output, -o — Output directory (defaults to skills/<skill-name>/)
• --verbose, -v — Show dataset analysis details

Example:

omniskill generate data/api-specs/ --name api-assistant --output skills/api-assistant/

create — Create Skill Scaffolding

Create an empty skill directory with template files:

omniskill create <skill-name> [--force]

Options:

• --force, -f — Overwrite existing skill directory

search — Search a Knowledge Base

omniskill search <query> --skill-dir <path> [options]

Options:

• --skill-dir, -d — Path to the skill directory (required)
• --format, -f — Output format: xml, markdown (default: xml)
• --limit, -l — Maximum number of results (default: 10)
• --type, -t — Filter by document type: csv, markdown
• --tag — Filter by tag (AND logic, repeatable)
• --metadata — Include BM25 scores in output
• --verbose, -v — Enable verbose output

Python API

Generate a Skill Programmatically

from omniskill.core.generator import generate_skill

analysis = generate_skill(
    dataset_dir="data/my-datasets",
    skill_name="my-skill",
    output_dir="skills/my-skill",
)

print(f"Generated {analysis.skill_name} with {analysis.total_documents} documents")

SearchEngine

Index and search directories of CSV/Markdown files:

from omniskill.core.engine import SearchEngine
from omniskill.core.assembler import OutputFormat, PromptAssembler

engine = SearchEngine()
engine.index_directory("skills/my-skill/datasets")

results = engine.search("API design", limit=10)

assembler = PromptAssembler()
print(assembler.assemble(results, output_format=OutputFormat.XML))

Dataset Analysis

Analyze a dataset directory without generating files:

from omniskill.core.generator import analyze_dataset

analysis = analyze_dataset("data/my-datasets")
print(f"CSV files: {len(analysis.csv_files)}")
print(f"Markdown files: {len(analysis.markdown_files)}")
print(f"Total documents: ~{analysis.total_documents}")

Output Formats

OmniSkill supports three output formats for assembled search results:

Format	CLI Flag	Description
XML	--format xml	Structured <context_injection> with <rules> and <reference> sections
Markdown	--format markdown	Human-readable sections with source attribution
llms.txt	(in generated scripts)	Follows the llms.txt spec for LLM consumption

Contributing

Development Setup

git clone https://github.com/longcipher/omni-skill.git
cd omni-skill
uv sync --all-groups

Common Commands

just format      # Format code
just lint        # Run linter
just test        # Run unit tests
just bdd         # Run BDD tests
just test-all    # Run all tests
just build       # Build package
just typecheck   # Run type checker

Adding New Features

1. Write a failing Gherkin scenario in features/*.feature
2. Write a failing pytest test for the inner domain logic
3. Implement the feature
4. Re-run just test and just bdd to verify

License

Apache-2.0