Gemini Gem Converter

Convert Agent Skills into Gemini Gems with awareness of platform constraints and optimal adaptation strategies.

Platform Constraints Summary

For detailed constraints and workarounds, see references/gem-constraints.md.

Conversion Workflow

Step 1: Audit the Source Skill

Inventory all files in the source skill directory:

1. •Read SKILL.md — note frontmatter fields and body length
2. •List all files in scripts/, references/, and assets/
3. •Count total files (Gems allow max 10 knowledge files)
4. •Identify scripts that need conversion to instructional docs

Step 2: Map SKILL.md to Gem Instructions

Transformation rules:

1. •Remove Claude-specific directives — Strip file path references (/mnt/skills/...), remove tool invocation syntax, replace "Claude" with neutral terms ("this assistant")
2. •Adapt progressive disclosure — Gems don't support lazy loading. Consolidate critical info into instructions. Use knowledge files for supplementary detail.
3. •Remove frontmatter — Gems don't use YAML frontmatter. Extract name and description into the Gem configuration UI.

Step 3: Convert Bundled Resources

Use this naming convention for the flat file structure:

Original:                           Derived:
references/api-docs.md         →    REF_api-docs.md
references/workflows/create.md →    REF_workflows_create.md
scripts/rotate_pdf.py          →    SCRIPT_rotate_pdf.md (converted)
assets/template.pptx           →    ASSET_template.pptx

Prefix system:

• •REF_ — Reference documentation
• •SCRIPT_ — Script logic (converted to readable format)
• •ASSET_ — Binary assets (images, templates)
• •WORKFLOW_ — Multi-step procedures

Step 4: Convert Scripts to Instructional Documents

Since Gems cannot execute code, transform each script into:

1. •Procedural documentation — Step-by-step instructions the AI can explain
2. •Code snippets for users — Code the user can copy/paste and run locally
3. •Guidance on when to use — Context for when this procedure applies

Example transformation:

Source (scripts/rotate_pdf.py):

def rotate_pdf(input_path, degrees):
    from pypdf import PdfReader, PdfWriter
    reader = PdfReader(input_path)
    writer = PdfWriter()
    for page in reader.pages:
        page.rotate(degrees)
        writer.add_page(page)
    return writer

Target (SCRIPT_rotate_pdf.md):

# PDF Rotation Procedure

## When to use
User wants to rotate pages in a PDF document.

## Steps to guide user
1. Explain that PDF rotation requires a Python environment
2. Provide this code for the user to run:
   [Python code snippet]
3. Offer to explain any part of the code
4. Suggest online alternatives if user cannot run Python

Step 5: Consolidate to 10-File Limit

Gems allow a maximum of 10 knowledge files. If the source skill has more than 10 resource files:

1. •Merge related references into single files (e.g., combine all API docs)
2. •Prioritize core documentation — drop files that are nice-to-have
3. •Inline short references into the instructions field
4. •Aim for 5-7 files to leave room for future additions

Step 6: Test the Gem

1. •Create the Gem in Google AI Studio with the converted instructions
2. •Upload all knowledge files
3. •Test with representative queries from the original skill's use cases
4. •Verify the Gem finds and uses knowledge files correctly
5. •Iterate on instructions if the Gem misses important context

Platform Opportunities

Google Drive integration: Gems can reference Google Docs/Sheets that auto-update. For skills with frequently changing reference data, consider linking a Google Sheet instead of a static knowledge file.

Naming Convention Quick Reference

SKILL.md fields → Gem Configuration:
  name          → Gem Name
  description   → Instructions (first paragraph)
  body          → Instructions (full content)

Resource files → Knowledge Files:
  references/*  → REF_*.md
  scripts/*     → SCRIPT_*.md (converted from code to docs)
  assets/*      → ASSET_* (uploaded directly if supported type)

Quality Expectations

Script-heavy skills will lose the most capability. Knowledge-heavy skills translate well.