Add Skill Creator skill (Create/Eval/Improve/Benchmark) #738
Description
Overview
Restructure and add Anthropic's skill-creator, released in March 2026, as a codingbuddy multi-tool skill.
Background
- Anthropic's skill-creator is a framework managing the entire skill development lifecycle
- 4 modes: Create, Eval, Improve, Benchmark
- 4 internal agents: Executor, Grader, Comparator, Analyzer
- March 2026 update added the eval pipeline (eval/improve/benchmark)
codingbuddy Differentiation
The Anthropic original is Claude Code-exclusive. The codingbuddy version:
| Aspect | Anthropic Original | codingbuddy Version |
|---|---|---|
| Target tools | Claude Code only | Multi-tool (Cursor, Claude Code, Codex, Q, Kiro, Antigravity) |
| Skill tool dependency | Claude Code Skill tool | Tool-specific skill loading guidance |
| CC-specific frontmatter | context: fork, agent, etc. |
Per-tool support/unsupport noted |
| Skill quality criteria | Anthropic standards | Reflects codingbuddy Iron Law, testable criterion, etc. |
| Implementation | Original copy | Self-rewritten (license-safe) |
Relationship with Existing Skills
| Existing Skill | Scope | Conflict? |
|---|---|---|
rule-authoring |
AI rule writing (multi-tool) | ❌ Rules ≠ Skills |
agent-design |
Agent JSON definition | ❌ Agents ≠ Skills |
prompt-engineering |
Prompt optimization |
Conclusion: Add as a new skill, no conflict with existing skills
Final Directory Structure
packages/rules/.ai-rules/skills/skill-creator/
├── SKILL.md # Main (4-mode workflow)
├── agents/
│ ├── grader.md # Grading agent
│ ├── analyzer.md # Analysis agent
│ └── comparator.md # Blind comparison agent
├── references/
│ ├── schemas.md # eval/benchmark JSON schema definitions
│ ├── frontmatter-guide.md # v2.0 frontmatter field guide
│ └── multi-tool-compat.md # Multi-tool compatibility matrix
├── scripts/
│ ├── aggregate_benchmark.py # Benchmark aggregation
│ ├── run_loop.py # Description optimization loop
│ └── init_skill.sh # Scaffolding
├── assets/
│ ├── eval_review.html # Trigger accuracy eval viewer
│ └── skill-template.md # SKILL.md template
└── eval-viewer/
└── generate_review.py # Result viewer generator
Sub-issues
Split into 5 sub-issues based on dependencies.
A (SKILL.md) ──→ B (agents/)
C (references/) ──→ D (scripts/)
──→ E (assets + eval-viewer/)
- [A] Write skill-creator SKILL.md main file #739 — [A] Write SKILL.md main file (4-mode workflow)
- [B] Write skill-creator agents/ agent instructions #740 — [B] Write agents/ agent instructions (after A is complete)
- [C] Write skill-creator references/ reference documents #741 — [C] Write references/ reference documents (parallelizable with A)
- [D] Implement skill-creator scripts/ #742 — [D] Implement scripts/ (after C is complete)
- [E] Implement skill-creator assets/ + eval-viewer/ viewers and templates #743 — [E] Implement assets/ + eval-viewer/ viewers and templates (after C is complete)
License
- No direct copying of Anthropic original code
- Self-rewritten to fit codingbuddy context
- Only JSON schema interfaces are referenced
- Attribution to Anthropic skill-creator added to README if needed