multi-ai-skill-builder（多 AI 技能构建者）

Multi-AI Skill Builder

Overview

multi-ai-skill-builder is a meta-skill that systematizes the Multi-AI approach to building Claude Code skills. It coordinates multiple AI models (Claude, Gemini, Codex) through research, planning, and implementation phases to create comprehensive, production-ready skills.

Purpose: Build high-quality Claude Code skills using Multi-AI research and synthesis

Pattern: Workflow-based (5-step sequential process)

Key Principles (validated by tri-AI research):

1. Multi-Source Research - Claude for docs, Gemini for web, Codex for GitHub
2. Synthesis Before Building - Combine findings into coherent plan
3. Progressive Disclosure - SKILL.md + references/ + scripts/
4. Pattern Compliance - Follow established skill patterns
5. Validation Loop - Multi-AI review of generated skills
6. Iterative Refinement - Build → Review → Improve cycle

Quality Targets:

• Research coverage: 3+ sources (Claude + Gemini + Codex)
• Skill completeness: All required sections present
• Code examples: 5+ practical examples per skill
• Validation score: ≥85/100

When to Use

Use multi-ai-skill-builder when:

• Creating new Claude Code skills from scratch
• Building skill families (related skills for a domain)
• Enhancing existing skills with new research
• Researching best practices for a technical domain
• Converting research into actionable skills
• Establishing skill development workflows

When NOT to Use:

• Simple skill updates (use direct editing)
• Trivial skills (<100 lines, single operation)
• Skills outside your domain expertise (research first)

Prerequisites

Required

• Skill topic/domain clearly defined
• Time for research (30-60 min) and building (60-120 min)
• Access to Claude (always available)

Recommended

• Gemini CLI for web research
• Codex CLI for GitHub patterns
• Existing example skills to reference

Understanding

• Claude Code skill structure
• YAML frontmatter format
• Progressive disclosure architecture

Workflow

Step 1: Research Phase (Multi-AI)

Time: 30-60 minutes Purpose: Gather comprehensive knowledge from multiple sources

1.1 Claude Documentation Research

Launch Claude subagent for official documentation:

Research [TOPIC] for Claude Code skill creation:

Focus on:
1. Official documentation and best practices
2. Existing similar skills in the codebase
3. API patterns and SDK usage
4. Common workflows and use cases

Output structured findings with:
- Key concepts
- Recommended patterns
- Code examples
- Gotchas and anti-patterns

1.2 Gemini Web Research

Use Gemini CLI for current best practices:

gemini -p "Research [TOPIC] best practices 2024-2025:
1. Industry standard approaches
2. Common patterns and anti-patterns
3. Tool comparisons and recommendations
4. Recent developments and trends
5. Real-world implementation examples

Provide comprehensive findings with sources."

1.3 Codex GitHub Research

Use Codex for code patterns:

codex "Research GitHub patterns for [TOPIC]:
1. Popular library implementations
2. Production code examples
3. Testing patterns
4. Configuration approaches
5. Error handling patterns

Provide code examples and best practices."

1.4 Create Research Directory

mkdir -p .analysis/[topic]-research

Save all research to:

• .analysis/[topic]-research/claude-docs-research.md
• .analysis/[topic]-research/gemini-web-research.md
• .analysis/[topic]-research/codex-github-research.md

Step 2: Synthesis Phase

Time: 15-30 minutes Purpose: Combine research into actionable plan

2.1 Synthesize Findings

Synthesize findings from multi-AI research:

Claude findings: [SUMMARY]
Gemini findings: [SUMMARY]
Codex findings: [SUMMARY]

Create unified synthesis:
1. Key patterns to implement
2. Best practices to follow
3. Anti-patterns to avoid
4. Recommended skill structure
5. Operations/workflows to include
6. Code examples to provide

2.2 Create Synthesis Document

Save to .analysis/[topic]-research/SYNTHESIS_AND_PLAN.md:

# [Topic] Skill Synthesis

## Research Sources
- Claude: Documentation analysis
- Gemini: Web best practices
- Codex: GitHub patterns

## Key Findings
1. [Finding 1]
2. [Finding 2]
...

## Recommended Structure
- Pattern: [workflow/task/reference/capabilities]
- Operations: [list]
- References: [list]

## Implementation Plan
1. Create SKILL.md with [structure]
2. Add references for [topics]
3. Include [N] code examples
4. Cover [operations/workflows]

## Quality Checklist
- [ ] YAML frontmatter complete
- [ ] Trigger keywords included
- [ ] 5+ code examples
- [ ] Error handling covered
- [ ] All patterns validated

Step 3: Build Phase

Time: 60-90 minutes Purpose: Create the skill files

3.1 Create Directory Structure

mkdir -p .claude/skills/[skill-name]/references
mkdir -p .claude/skills/[skill-name]/scripts  # if needed

3.2 Build SKILL.md

Follow the template structure:

---
name: skill-name-in-hyphen-case
description: [Purpose]. [Pattern type]. Use when [triggers].
allowed-tools: Task, Read, Write, Edit, Glob, Grep, Bash
---

# Skill Name

## Overview
[Brief description]
**Purpose**: [One line]
**Pattern**: [Workflow/Task/Reference/Capabilities]
**Key Principles**: [3-6 numbered principles]
**Quality Targets**: [Measurable goals]

## When to Use
[Use cases and non-use cases]

## Prerequisites
### Required / ### Recommended / ### Understanding

## [Operations or Workflow Steps]
[Main content with code examples]

## Multi-AI Coordination
[How to use Claude/Gemini/Codex for this skill]

## Related Skills
[Links to related skills]

## References
[Links to reference files]

3.3 Build Reference Files

Create detailed guides in references/:

• Detailed how-to guides
• Configuration references
• Integration patterns
• Troubleshooting guides

3.4 Add Code Examples

Every skill needs:

• Quick start example
• Common use case examples
• Advanced/edge case examples
• Error handling examples
• Integration examples

Step 4: Validation Phase

Time: 15-30 minutes Purpose: Verify skill quality

4.1 Structure Validation

Check YAML frontmatter:

head -20 .claude/skills/[skill-name]/SKILL.md

Verify sections:

• YAML frontmatter with name, description
• Overview section
• When to Use section
• Prerequisites section
• Main content (operations/workflows)
• Related Skills section
• References section

4.2 Multi-AI Review

Review this skill for quality:

[PASTE SKILL.md]

Check:
1. YAML frontmatter complete and descriptive?
2. Trigger keywords in description?
3. Clear when to use / when not to use?
4. Prerequisites documented?
5. 5+ code examples?
6. Error handling covered?
7. Progressive disclosure followed?
8. Related skills linked?

Score (0-100) and improvement suggestions.

4.3 Gemini Cross-Check

gemini -p "Verify this skill against best practices:
[SKILL CONTENT]

Check for:
- Accuracy of technical information
- Missing important patterns
- Outdated recommendations"

Step 5: Refinement Phase

Time: 15-30 minutes Purpose: Apply improvements from validation

5.1 Apply Feedback

Address issues from validation:

• Fix any structural issues
• Add missing examples
• Clarify unclear sections
• Enhance descriptions

5.2 Final Quality Check

Ensure:

• Score ≥85/100
• All validation items pass
• Cross-check feedback addressed
• Ready for production use

5.3 Create Delivery Summary

Save to .analysis/[topic]-research/DELIVERY_SUMMARY.md:

# [Skill Name] - Delivery Summary

**Date**: [Date]
**Status**: COMPLETE
**Total Lines**: [X] lines across [Y] files

## Research Phase
- Claude: [Summary]
- Gemini: [Summary]
- Codex: [Summary]

## Skills Delivered
### [skill-name]
**Files**:
- SKILL.md ([X] bytes)
- references/[file].md
...

**Coverage**:
- [Feature 1]
- [Feature 2]
...

## Quality Validation
- Structure: PASS
- Content: PASS
- Examples: [N] included
- Score: [X]/100

## Usage Examples
[Show example triggers]

Multi-AI Coordination

Agent Assignment for Skill Building

Research Commands

Claude Subagent:

Use Task tool with subagent_type=Explore

Gemini CLI:

gemini -p "Research [topic]: [specific questions]"

Codex CLI:

codex "Research GitHub patterns for [topic]"

Templates

SKILL.md Template

See templates/SKILL_TEMPLATE.md

Reference File Template

See templates/REFERENCE_TEMPLATE.md

Synthesis Template

See templates/SYNTHESIS_TEMPLATE.md

Quality Checklist

Structure (20 points)

• YAML frontmatter valid
• All required sections present
• Progressive disclosure followed
• File naming conventions

Content (25 points)

• Clear descriptions
• Comprehensive coverage
• Accurate information
• Well-organized

Examples (25 points)

• 5+ code examples
• Quick start example
• Common use cases
• Error handling
• Advanced scenarios

Usability (15 points)

• Easy to navigate
• Clear when to use
• Prerequisites documented
• Related skills linked

Validation (15 points)

• Multi-AI reviewed
• Cross-checked
• Feedback addressed
• Score ≥85/100

Example: Building ECS Skills

This example shows how we built the ECS/Fargate skill family:

Research Phase

# Claude subagent for AWS docs
# Gemini for 2024-2025 best practices
gemini -p "Research ECS/Fargate best practices 2024-2025..."

# Codex for GitHub patterns
codex "Research GitHub patterns for ECS/Terraform..."

Synthesis Phase

• Combined findings into unified plan
• Identified 5 skills to build
• Mapped to existing EKS skill patterns

Build Phase

• Created boto3-ecs (SDK patterns)
• Created terraform-ecs (IaC)
• Created ecs-fargate (Fargate specifics)
• Created ecs-deployment (strategies)
• Created ecs-troubleshooting (debugging)

Result

• 2,209+ lines across 15 files
• All skills validated
• Progressive disclosure implemented
• Multi-AI researched and reviewed

Related Skills

• multi-ai-research: Research phase patterns
• multi-ai-planning: Planning phase patterns
• multi-ai-implementation: Implementation patterns
• multi-ai-verification: Validation patterns
• skill-builder-generic: Universal skill patterns
• review-multi: Skill review framework

References

• templates/SKILL_TEMPLATE.md - Skill file template
• templates/REFERENCE_TEMPLATE.md - Reference file template
• templates/SYNTHESIS_TEMPLATE.md - Research synthesis template