Create M4 Skill

M4 Skill Creator

This skill guides you through creating properly formatted M4 skills with complete provenance tracking and appropriate tier and category assignment.

When to Use This Skill

• User asks to create a new M4 skill
• User wants to document a clinical scoring system, cohort definition, or research pattern
• User wants to document an M4 framework feature or workflow
• User mentions "new skill", "create skill", "document [concept]"
• User wants to contribute to the M4 skills library

Interactive Creation Process

Follow these steps in order. At each step, gather information from the contributor and validate completeness before proceeding.

Step 1: Understanding the Skill

Interview the contributor with these questions:

1. What concept does this skill address?
2. Is this a clinical skill (scoring system, cohort definition, lab interpretation, medication calculation) or a system skill (M4 framework feature, data structure, workflow pattern)?
3. When should an AI assistant activate this skill? (Specific triggers)
4. Is this based on:
   • Published source code (e.g., MIT-LCP mimic-code)?
   • Original domain expertise?
   • Community contribution?
5. What are the key references? (Publications for clinical, documentation/repos for system)

Do NOT auto-generate answers. Wait for contributor input at each question.

Step 2: Determine Category and Tier

Based on their answers, assign category and guide tier selection:

Category (from question 2):

• clinical — Encodes clinical domain knowledge
• system — Encodes M4 framework knowledge, data structure guidance, or meta-skills

Tier selection for clinical skills:

• validated — Derives from published, validated source (mimic-code, peer-reviewed paper). Has URL to source. Will be reviewed by at least one clinician for clinical accuracy.
• expert — Original work by a clinician or domain expert. Will be reviewed by at least one independent clinician. Creator has clinical expertise.
• community — Contribution without required clinical reviews. Technical review for correctness.

Tier selection for system skills:

• validated — Based on documented, stable APIs or established patterns. Will be reviewed for technical accuracy.
• expert — Original work by someone with relevant technical expertise. Will be reviewed by at least one independent technical reviewer.
• community — Contribution with technical review for correctness.

Ask: "Based on your answers, this appears to be a [category] skill at [tier] tier. Does that sound right? Who will review it?"

Step 3: Create Frontmatter

Construct the YAML header using exactly these four fields:

---
name: kebab-case-name
description: One-sentence description stating what the skill does and when to use it.
tier: [validated|expert|community]
category: [clinical|system]
---

Guidelines:

• name: Match directory name, use kebab-case
• description: Should trigger skill activation. Include "Use for..." or "Use when..." context.
• tier: From Step 2
• category: From Step 2
• No other fields in the frontmatter. Licensing, authorship, and source URLs go in PROVENANCE.yaml.

Step 4: Structure the Body

Guide them through these required sections:

# [Skill Title]

Brief introduction (1-2 sentences).

## When to Use This Skill

- [Bullet list of activation scenarios]
- [Be specific about contexts]

## [Core Content Sections]

[Tables, thresholds, scoring criteria, API reference,
workflow steps — whatever is needed]

## Critical Implementation Notes

[Gotchas, edge cases, common mistakes to avoid]

## Example Queries

[Working SQL or code examples with logic explained in comments]

## References

- [Author et al. "Title." Journal. Year;Volume:Pages.]

Interview questions for clinical skills:

• "What are the core components or criteria?"
• "What mistakes do people commonly make?"
• "Can you provide a working SQL example?"
• "What are the 2-3 key publications?"

Interview questions for system skills:

• "What is the primary workflow or API surface?"
• "What are the common pitfalls or misuse patterns?"
• "Can you provide a working code example?"
• "What documentation or repos should be referenced?"

Dataset-agnostic design: Keep the body focused on the concept, not dataset-specific implementations. If dataset-specific SQL is needed, put it in separate script files (scripts/mimic-iv.sql, scripts/eicu.sql) rather than embedding multiple versions in the body.

Important: Don't add content the contributor hasn't provided. Ask questions to fill gaps.

Step 5: Create PROVENANCE.yaml

Build provenance file with contributor input:

sources:
  - url: [Ask: "What's the source URL?"]
    description: [Ask: "What does this source provide?"]
    license: [Ask: "What license?" Common: Apache-2.0, CC-BY-4.0, MIT]

created:
  by: [Ask: "Your name and credentials?" Format: "Jane Smith, MD"]
  role: [authoring|clinical-extraction|adaptation]
  date: [Use today's date: YYYY-MM-DD]

reviews:
  - by: [Ask: "Who will review this?" Format: "John Doe, MD"]
    date: [Leave TBD or set future date]
    scope: [clinical-accuracy for clinical skills, technical-accuracy for system skills]
    notes: [Optional - ask if they want to add notes]

references:
  - [Same as SKILL.md references - ask them to confirm]

changelog:
  - version: "1.0"
    date: [Today's date]
    summary: Initial skill creation

Special cases:

• No sources section if original work with no external source
• reviews required for validated and expert tiers
• reviews can be TBD with future dates if review is pending

Step 6: Validation Checklist

Before finalizing, verify with the contributor:

SKILL.md:

• Valid YAML frontmatter with exactly four fields: name, description, tier, category
• Name matches directory name (kebab-case)
• Description includes "Use for..." or "Use when..." triggers
• "When to Use This Skill" section present
• At least one working example (SQL or code)
• References section with key publications or documentation
• Body is dataset-agnostic (dataset-specific SQL in scripts/ if needed)
• No over-engineering (only what's needed for the AI to use the skill correctly)

PROVENANCE.yaml:

• Creator name and credentials included
• Role matches skill origin
• Reviews listed (with at least one for validated/expert)
• Sources documented (if applicable)
• Changelog has initial entry

Ask: "I've checked the structure. Would you like me to create the files now, or would you like to review the content first?"

Implementation Notes

1. Create directory structure:

   mkdir -p src/m4/skills/[skill-name]
   

2. Write files:

   • Write SKILL.md with frontmatter and body
   • Write PROVENANCE.yaml with complete provenance

3. If SQL scripts exist:

   mkdir -p src/m4/skills/[skill-name]/scripts
   

   Use separate files per dataset: mimic-iv.sql, eicu.sql.

4. Verify files are created: Read back both files to confirm they exist and are well-formed.

5. Update SKILLS_INDEX.md: Add the new skill to the appropriate category table.

Quality Guidelines

Context efficiency:

• Only include information the AI doesn't already know
• Be precise about edge cases
• Keep skills focused and scoped

Dataset-agnostic design:

• Document the concept, not the dataset
• SQL in the body should work across datasets or use generic table references
• Dataset-specific implementations go in scripts/

Maintainability:

• Clear provenance enables future updates
• Examples serve as regression tests
• References enable verification

Anti-Patterns to Avoid

• Don't auto-generate content: Wait for contributor input at every step
• Don't skip provenance: Required for scientific accountability
• Don't add boilerplate: Only include what's needed for the AI
• Don't guess tier or category: Interview to determine proper assignment
• Don't skip reviews: Validated/expert tiers require reviews
• Don't embed dataset-specific SQL in the body: Use the scripts/ directory
• Don't add extra frontmatter fields: Only name, description, tier, category

Reference

For the canonical format specification, see src/m4/skills/SKILL_FORMAT.md.