Skill Architect

skill-architect

Build, validate, and improve Claude Code skills using a design-first methodology.

how to use

• /skill-architect — apply skill architecture principles to the current conversation.
• /skill-architect build <name> — create a new skill from scratch following the full build workflow.
• /skill-architect review <path> — review an existing skill and produce a scored quality report.
• /skill-architect validate <path> — run the validator and report structural/frontmatter/content issues.

when to apply

• Creating a new Claude Code skill from scratch.
• Improving or refactoring an existing skill.
• Auditing skill quality across a skill library.
• Troubleshooting why a skill is not triggering correctly.
• Reviewing a skill before publishing or sharing.
• Converting a one-off workflow into a reusable skill.

when NOT to apply

• General coding tasks — use coding-workflow instead.
• "Improve my skills" meaning career or personal development.
• Prompt engineering for chat conversations (skills are file-based, not chat-based).
• Writing one-off documentation or READMEs.
• Configuring MCP servers (that is infrastructure, not skill authoring).
• Iterative skill improvement with eval benchmarks — use skill-creator instead; it provides subagent testing, programmatic grading, eval-viewer, and description optimization with eval queries.

rule priorities

Full per-category rules: references/skill-system-patterns.md.

critical constraints

Design-first (priority 1)

• MUST define 3+ trigger prompts before writing.
• MUST identify the skill category (document/asset, workflow automation, MCP enhancement).
• MUST check existing skills for overlap.
• MUST define explicit scope boundaries — what the skill does NOT cover.
• NEVER start writing SKILL.md without completing the design phase.

YAML frontmatter (priority 2)

• MUST start with --- on line 1.
• MUST include name in kebab-case matching the folder name exactly.
• MUST include description under 1024 characters with no angle brackets.
• MUST NOT use forbidden terms in name: claude, anthropic, plugin, extension.
• The description's first sentence determines triggering.
• Use > folded scalar for multi-line descriptions.

Full field spec: references/yaml-reference.md.

Folder structure (priority 3)

skill-name/
├── SKILL.md          # Required
├── scripts/          # Optional — stdlib only
└── references/       # Optional — supplementary docs

• Folder name MUST be kebab-case and MUST match the name field.
• MUST NOT have both README.md and SKILL.md.
• Scripts MUST use Python stdlib only — no pip dependencies.
• Keep SKILL.md under 5000 words; spill into references/.

core workflow

Build (creating a new skill)

1. Design — define triggers, category, scope, and name.
2. Structure — create folder layout.
3. Content — write scripts → references → SKILL.md.
4. Validate — run scripts/validate_skill.py (13 checks must pass).
5. Test — Tier 1 triggering, Tier 2 functional.
6. Refine — register, iterate on real usage.

Full procedure with templates: references/architecture-workflow.md.

Review (auditing an existing skill)

1. Read SKILL.md, references/, scripts/.
2. Validate with validate_skill.py --verbose.
3. Score 8 dimensions (0-5 each, 40 max).
4. Output validator results, scorecard, top 3 issues, fixes.

Full scoring rubric and review procedure: references/evaluation-and-validation.md.

skill categories

Per-category patterns: references/skill-system-patterns.md.

validation gates

Structural validation — all 13 checks must pass before shipping:

python3 ~/.claude/skills/skill-architect/scripts/validate_skill.py <skill-path>

Testing tiers — full methodology in references/testing-guide.md:

safety constraints

• Do not create misleading skills.
• Do not include malware, exploit code, credential harvesting, or data exfiltration patterns.
• Do not bundle secrets, credentials, tokens, private context, personal identifiers, or local user paths in skill files.
• Preserve existing name and frontmatter identity when improving an installed skill unless the user explicitly asks for a rename.
• Treat bundled scripts as code: keep them scoped, inspectable, testable, and stdlib-only.

output expectations

When creating or modifying a skill, report:

• Skill folder and files created or changed.
• Trigger behavior (positive + negative scope).
• Safety constraints included.
• Validator results (which of 13 checks passed/failed).
• Tests performed and results.
• Known limitations and recommended next iteration.

minimal frontmatter example

---
name: example-skill
description: >
  One-sentence trigger statement. Use when <specific user intent A> or
  <specific user intent B>.
version: 1.0.0
keywords: domain, capability, target
---

Full templates and patterns: references/yaml-reference.md, references/workflow-patterns.md.

Reference map