Skill Production Pipeline — Claude-skills

Skill Production Pipeline — claude-skills

**Effective: 2026-03-07** | Applies to ALL new skills, improvements, and deployments. **Owner:** Leo (orchestrator) + Reza (final approval)

---

Mandatory Pipeline

Every skill MUST go through this pipeline. No exceptions.

Intent → Research → Draft → Eval → Iterate → Compliance → Package → Deploy → Verify → Rollback-Ready

Tool: Anthropic Skill Creator (v2025-03+)

**Location:** `~/.openclaw/workspace/skills/skill-creator/` **Components:** SKILL.md, 3 agents (grader, comparator, analyzer), 10 scripts, eval-viewer, schemas

Dependencies

Tool	Version	Install	Fallback
Tessl CLI	v0.70.0	tessl login (auth: rezarezvani)	Manual 8-point compliance check
ClawHub CLI	latest	npm i -g @openclaw/clawhub	Skip OpenClaw publish, do manually later
Claude Code	2.1+	Already installed	Required, no fallback
Python	3.10+	System	Required for scripts

Iteration Limits

• Max 5 iterations per skill before escalation
• Max 3 hours per skill in eval loop
• If stuck → log issue, move to next skill, revisit in next batch

---

Phase 1: Intent & Research

1. Capture intent — What should this skill enable? When should it trigger? Expected output format?
2. Interview — Edge cases, input/output formats, success criteria, dependencies
3. Research — Check competing skills, market gaps, related domain standards
4. Define domain expertise level — Skills must be POWERFUL tier (expert-level, not generic)

Phase 2: Draft SKILL.md

Using Anthropic's skill-creator workflow:

Required Structure

skill-name/
├── SKILL.md              # Core instructions (YAML frontmatter required)
│   ├── name: (kebab-case)
│   ├── description: (pushy triggers, when-to-use)
│   └── Body (<500 lines ideal)
├── scripts/              # Python CLI tools (no ML/LLM calls, stdlib only)
├── references/           # Expert knowledge bases (loaded on demand)
├── assets/               # Templates, sample data, expected outputs
├── agents/               # Sub-agent definitions (if applicable)
├── commands/             # Slash commands (if applicable)
└── evals/
    └── evals.json        # Test cases + assertions

SKILL.md Rules

• YAML frontmatter: name + description required
• Description must be "pushy" — include trigger phrases, edge cases, competing contexts
• Under 500 lines; overflow → reference files with clear pointers
• Explain WHY, not just WHAT — theory of mind over rigid MUSTs
• Include examples with Input/Output patterns
• Define output format explicitly

Phase 3: Eval & Benchmark

3a. Create Test Cases

• 2-3 realistic test prompts (what real users would actually say)
• Save to evals/evals.json (schema: references/schemas.md)
• Include files for file-dependent skills

3b. Run Evals

• Spawn with-skill AND baseline (without-skill) runs in parallel
• Save to `-workspace/iter