Workflow Skill Creator

When to Use

Use when you need to:

• Convert a completed analysis workflow into a reusable skill
• Package a successful analysis session for future reuse
• Create a new skill from scratch following best practices

Do NOT use when you need to:

• Run an existing skill → use the specific skill directly
• Fix a broken skill → edit the skill files directly
• Understand skill architecture → read the existing skills as examples

Overview

This meta-skill guides the creation of new analysis skills following the Rhine Life Lab skill format. It enforces a structured 4-phase process to ensure high-quality, reusable skills.

Phase 1: Understand the Workflow (MANDATORY)

Before writing any code, answer these questions:

1. What is the input? (h5ad format? what preprocessing state?)
2. What is the output? (h5ad? CSV? plots? JSON results?)
3. What are the key steps? (list each analysis step)
4. What dependencies are needed? (scanpy? squidpy? scipy?)
5. What can go wrong? (edge cases, missing data, convergence failures)
6. What other skills does this depend on? (prerequisite skills)
7. What skills does this feed into? (downstream skills)

Phase 2: Skill Design

Create the directory structure:

~/.agents/skills/<skill-name>/
  SKILL.md              # Skill definition
  scripts/
    <script>.py         # CLI wrapper script

SKILL.md Template

---
name: <kebab-case-name>
description: "<trigger phrase starting with 'Use when'>"
---

Required sections (in order):

1. # Title
2. ## Prerequisites
3. ## When to Use (with "Do NOT use" subsection and cross-references)
4. ## Core Rules
5. ## Workflow (with Step N subsections)
6. ## Common Mistakes
7. ## Troubleshooting (table format)
8. ## Output (JSON schema)

Cross-Reference Format

Use bold skill names in parentheses:

• → (use **scrna-qc-preprocess**) for prerequisites
• → (use **scrna-clustering-umap**) for downstream skills

Phase 3: Implementation

Rules

1. Reuse scrnatools: Import from ~/.agents/skills/_shared/scrnatools/ for I/O, QC, stats, output
2. Wrapper script pattern: Use argparse with subcommands, --input/--output required
3. File output: All results go to JSON file via SkillOutput, stdout for status only
4. Error handling: Use try/except with fallback for every optional dependency
5. Sparse-first: Always check sparse.issparse(X) and prefer sparse operations
6. NaN handling: Call clean_nan() before any computation

Wrapper Script Template

#!/usr/bin/env python3
"""<Skill description>."""
import argparse, sys, os
sys.path.insert(0, os.path.join(os.path.dirname(__file__), "../../_shared"))
from scrnatools.io import load_h5ad, save_h5ad
from scrnatools.output import SkillOutput

def run(input_path, output_path, **kwargs):
    output = SkillOutput("<skill-name>", output_path)
    adata = load_h5ad(input_path)
    # ... analysis steps ...
    output.finalize(status="success")

def main():
    parser = argparse.ArgumentParser()
    sub = parser.add_subparsers(dest="command")
    p = sub.add_parser("run")
    p.add_argument("--input", required=True)
    p.add_argument("--output", required=True)
    args = parser.parse_args()
    if args.command == "run":
        run(args.input, args.output)

if __name__ == "__main__":
    main()

Phase 4: Validation

1. Test with real data (at least one h5ad file)
2. Verify JSON output is valid and complete
3. Check all cross-references point to existing skills
4. Ensure error messages are clear and actionable