BioMed Skill Creator

A meta-skill for creating and improving skills in the OpenBioMed biomedical toolkit.

Overview

This skill guides you through creating biomedical skills with interactive validation. Biomedical workflows require domain-specific validation with real data before finalization.

Workflow:

1. Capture Intent → 2. Design Workflow → 3. Interactive Validation → 4. Finalize → 5. Evaluate
        ↑                                                              ↓
        └──────────────────── Iterate if needed ←─────────────────────┘

Step 1: Capture Intent

Ask clarifying questions:

1. What biomedical task should this skill perform?
2. What inputs will users provide? (molecule name/SMILES, protein ID, text)
3. What outputs should the skill produce? (reports, files, predictions, visualizations)
4. Are there edge cases or constraints?

Input Types

Input Type	Factory Method	Example
Molecule	Molecule.from_smiles()	"CC(=O)OC1=CC=CC=C1C(=O)O"
Protein	Protein.from_fasta()	"MKFLILLFNILCLFPVLAADNH..."
Pocket	Pocket.from_protein_ref_ligand()	Protein + reference ligand
Text	Text.from_str()	"What is this molecule?"

Step 2: Design Workflow

Identify tools and steps. See references/tools_reference.md for available tools.

Common Workflow Patterns

Pattern	Tools Flow
Drug-likeness	molecule_name_request → molecule_qed/sa/logp/lipinski → summarize
Protein Mutation	protein_uniprot_request → mutation_explanation → protein_folding → visualize
Structure-Based Design	protein_pdb_request → extract_molecules → structure_based_drug_design → docking
Molecule Q&A	molecule_name_request → molecule_question_answering → format

Basic Workflow Code Pattern

from open_biomed.tools.tool_registry import TOOLS

# Get entity
tool = TOOLS["tool_name"]
result, message = tool.run(parameter=value)
entity = result.get("protein") or result.get("molecule")

# Process with other tools
another_tool = TOOLS["another_tool"]
output, msg = another_tool.run(entity=entity)

Step 3: Interactive Validation (CRITICAL)

Execute ONE step at a time and check with user before proceeding.

After designing the workflow, ask:

"Please provide an example input and I'll run through each step showing results."

For Each Step

1. Execute the step using OpenBioMed tools
2. Display results with standardized format (see references/validation_template.md)
3. Ask for feedback: "Is this result satisfactory? (yes/proceed/modify/skip)"

Handling Errors

When a step fails:

1. Explain the error clearly
2. Propose alternatives (fallback tools, web search, skip)
3. Ask user to decide

After All Steps

Present summary and ask:

"Do you want to:

1. Proceed with this workflow?
2. Modify and re-validate?
3. Try different input?"

Step 4: Finalize the Skill

Once approved, create the skill files:

Directory Structure

skill-name/
├── SKILL.md              # Main skill definition (< 200 lines)
├── examples/             # Runnable example scripts
│   └── basic_example.py
└── references/           # Detailed documentation
    ├── advanced.md
    └── troubleshooting.md

SKILL.md Template

See references/skill_template.md for the full structure. Key sections:

---
name: skill-name
description: >
  [One-line summary of what the skill does].
  Use this skill when:
  (1) [Use case 1],
  (2) [Use case 2],
  (3) [Use case 3].
license: [MIT|Apache-2.0|BSD-3-Clause|GPL-3.0]
category: [category from list below]
tags: [tag1, tag2, tag3]
---

# Skill Title

## When to Use
## Workflow (keep code snippets < 20 lines)
## Expected Outputs
## Error Handling

License Selection

Before finalizing SKILL.md, ask the user to choose a license:

"What license should this skill use?

1. MIT (Recommended) - Permissive, allows commercial use
2. Apache-2.0 - Permissive with patent grant
3. BSD-3-Clause - Permissive, no endorsement clause
4. GPL-3.0 - Copyleft, derivatives must be open source"

Default to MIT if user doesn't specify.

Category Options

Category	Description
drug-discovery	Drug design, molecule generation, lead optimization, virtual screening
admet-prediction	Absorption, distribution, metabolism, excretion, toxicity prediction
protein-engineering	Protein design, stability optimization, function prediction
protein-structure	Structure prediction, folding, conformational analysis
mutation-analysis	Mutation effect prediction, variant annotation, engineering
antibody-design	Antibody/nanobody design, affinity maturation, epitope prediction
immunology	Immunogenicity prediction, vaccine design, immune profiling
single-cell	Single-cell analysis, cell annotation, spatial transcriptomics
genomics	Gene analysis, variant calling, regulatory element prediction
transcriptomics	RNA-seq analysis, expression profiling, differential expression
metabolomics	Metabolite identification, pathway analysis, metabolic modeling
proteomics	Protein identification, PTM analysis, protein-protein interactions
pathway-analysis	Pathway enrichment, network analysis, systems biology
bioactivity-prediction	Activity prediction, target identification, bioassay analysis
binding-affinity	Docking, binding prediction, protein-ligand interactions
molecular-dynamics	MD simulation, conformational sampling, free energy calculation
chemical-synthesis	Retrosynthesis, reaction prediction, synthesis planning
safety-toxicology	Toxicity prediction, safety assessment, off-target effects
clinical-translational	Biomarker discovery, patient stratification, drug repurposing
bioimaging	Medical imaging analysis, cell segmentation, image-based profiling
knowledge-retrieval	Literature mining, database queries, knowledge graphs
multi-modal-reasoning	Cross-modal tasks, text-based molecule/protein tasks, QA
visualization	Molecular visualization, structure rendering, report generation
utilities	Meta-skills, workflow automation, helper tools, evaluation

Writing Guidelines

1. Keep SKILL.md under 200 lines - Move long code to examples/ or references/
2. Code snippets < 20 lines - Link to full examples
3. Include interpretation - What do scores/outputs mean?
4. Handle errors - What if tools/APIs fail?

Step 5: Evaluate the Skill

Run evaluation to ensure quality. See references/evaluation_reference.md for details.

1. Create 2-3 test cases with realistic prompts
2. Run grader - Compare with-skill vs baseline agents
3. Analyze results - Identify patterns and issues
4. Iterate if needed

Quick Reference

See references/quick_reference.md for:

• Workflow patterns summary
• Input type reference
• Score interpretation tables
• Evaluation checklist

Communication Style

Adapt to user's familiarity:

• Expert: Use technical terms (ADMET, TPSA, RMSD)
• Intermediate: Brief explanations
• Beginner: Analogies, explain why metrics matter

Checklist

Before finalizing:

• Workflow validated with real input
• User approved the workflow
• SKILL.md under 200 lines
• Long code in examples/
• Error handling documented
• Test cases created and graded