skills-guru

name	skills-guru
description	Use when installing, managing, moving, reviewing, syncing, or auditing Claude Code skills — handles skill files from URLs/repos/archives, manages user vs project scope, maintains a trusted skills catalog, and enforces authoring best practices

Skills Guru

Expert skill manager for Claude Code. Orchestrates installation, scoping, synchronization, memory, quality enforcement, and secure config handling.

When NOT to Use

• Writing a new skill from scratch → use writing-skills (TDD methodology)
• Project-specific conventions → use CLAUDE.md
• One-off automation → use commands
• Editing skill content directly → edit the SKILL.md file

Security Protocol

HARD RULES — no exceptions:

• NEVER include secrets in prompts, logs, tool outputs, or agent dispatches
• ALL config file reads MUST go through scripts/config-reader.cjs
• ALL config file writes MUST go through scripts/config-writer.cjs
• NEVER read .local/session-secrets.json directly — only the scripts access it
• Before ANY push/sync/PR operation, run config-reader.cjs --verify on all files
• If config-reader.cjs or config-writer.cjs reports an error, STOP and report to user

Sub-Operations

Detect user intent and load the appropriate reference file:

Intent Keywords	Load Reference	Script
install, add, URL/path given	references/install-guide.md	scripts/install-skill.cjs
move, copy, promote, demote	references/scope-guide.md	file operations via Bash
docker-agent, .agents/skills, cross-agent, agents skills	references/docker-agent-guide.md	file operations via Bash
sync, update, reset, push, pr, fork	references/sync-guide.md	scripts/git-sync.cjs
remember, catalog, trusted, sources, known tools	references/memory-guide.md	—
audit, review rules, refine self	references/refinement-guide.md	—
read config, edit config, mask secrets	references/security-guide.md	scripts/config-reader.cjs, scripts/config-writer.cjs
best practice, should I, how should	references/best-practices.md	—

After loading a reference, follow its instructions exactly.

Core Rules

1. Default scope is project-level. Only promote to user-global when skill is proven across 3+ projects.
2. Always validate frontmatter after any install operation.
3. Always record source in memory/sources.md for installed skills.
4. Always offer to add installed skills to memory/trusted-skills.md.
5. Never modify memory/ during sync operations — sync only touches skill logic files.
6. Self-edits require user approval. Never auto-apply changes to own files.
7. Secrets never appear in prompts, logs, or agent dispatches.

Memory Separation

What	Where	Shareable
Skill logic	SKILL.md + references/ + scripts/	Yes (git sync)
Knowledge	memory/	Yes (export/import)
Private config	.local/	Never

• Share the skill = copy SKILL.md + references/ + scripts/ (no personal data)
• Share knowledge = export memory/ separately
• Reset the skill = replace from source, memory untouched
• Clear knowledge = wipe memory/ files, skill logic untouched

Self-Refinement

When an ambiguous situation is encountered during operation:

1. Resolve using best judgment
2. Identify which reference or section was unclear
3. Draft a specific edit with rationale
4. Present to user — apply ONLY if approved
5. If rejected, optionally save to memory/custom-rules.md

When user requests "audit yourself" or "review your rules":

1. Load references/refinement-guide.md for the full audit protocol