Skill Creation Architect for Openclaw

1. Copy or place the skill directory in the agent's configured Skills root. The directory name must match the lowercase, hyphenated frontmatter name.
2. Ensure the package contains SKILL.md and, when applicable, references/, scripts/, assets/, evals/, and .skillignore.
3. If the skill includes executable scripts, verify a shebang, set -euo pipefail where appropriate, safe path handling, bounded iteration, and documented dependencies.
4. Set SKILL_BASE only when the Skills root cannot be discovered automatically.

export SKILL_BASE="/path/to/skills"

5. Run the script-checkable portion of the audit from the skill directory:

bash scripts/skill-audit.sh <skill-name>

6. Treat the script output as evidence, not as a complete audit result. Manually confirm all 55 checks, including cross-section semantics, security boundaries, qualitative content, degradation paths, and ICE-5 closure requirements.
7. For applicable skills, manually execute closure checks and evaluation tests. Do not deliver until the security scan passes and the responsible owner provides an explicit delivery confirmation.
8. Use the separate publishing or marketplace tool for installation, updates, or publication; this skill intentionally does not perform those actions.

Core package layout

Documentation taxonomy

• L1: Trigger words, use conditions, skip conditions, and non-applicable scenarios; loaded continuously and kept concise.
• L2: Numbered workflow, principles, gates, quick checks, child-agent rules, and actionable constraints; normally no more than 300 lines.
• L3: Detailed references loaded on demand, including schemas, examples, troubleshooting, AP-1 through AP-38 explanations, and evaluation guidance.
• P/D/E limitations: Known limitations are classified as Platform, Design, or Edge. Each entry records the capability boundary, trigger condition, and downgrade or human-review path; related limitations are merged and the list is capped at three prioritized hard limitations.
• Progress state: Long workflows externalize state in progress.json rather than relying on conversation context. Iteration controls use measurable fields such as stale_count, retry counts, timeouts, and completion evidence.
• Audit evidence: The 10 audit layers contain 55 total checks, with 29 script-checkable items and the remainder requiring human judgment and supporting evidence.