skills-manager

name	skills-manager
description	Full lifecycle management of LLM skills across the workstation — finding, archiving, installing, updating, and importing skills with their complete bundles (sub-skills + companion commands). Use when the user invokes /find-skills, /sync-skill, /install-skill, /update-skill, /import-skill, /push-skill, /search-skill, /audit-skills, or /backfill-diagrams.

Skills Manager

You are the domain expert for the LLM Skills Archive at C:\development\llm_skills. You understand the full lifecycle of skills: discovery, archiving, installation, updating, and importing project-level changes back to the archive.

Prompt the user for confirmation before any bulk or destructive operation. Never assume scope — always confirm what will change.

---

Core Concepts

Archive

C:\development\llm_skills is the single source of truth for all LLM skills, agents, and commands. It consolidates skills from the global Claude profile (~/.claude/) and all projects under C:\development\.

Skill Bundle

A skill directory may contain optional subdirectories that travel with it:

claude/skills/<name>/
├── SKILL.md              # Main skill
├── sub-skills/           # Thin delegate skills that extend this skill
│   └── <sub-name>/SKILL.md
└── commands/             # Slash-command dispatchers for this skill system
    └── <cmd>.md

Deploying a bundle to a repo copies:

• SKILL.md → <target>/.claude/skills/<name>/SKILL.md
• Each sub-skills/<sub>/SKILL.md → <target>/.claude/skills/<sub>/SKILL.md
• Each commands/<cmd>.md → <target>/.claude/commands/<cmd>.md

Installed Copy

A skill file with installed-from: llm_skills in its YAML frontmatter is an installed copy from the archive. The sync and find operations skip these — they are not project-developed skills.

Source Priority (conflict resolution)

When the same skill name exists in multiple sources, the highest-priority source wins:

1. Global profile (~/.claude/)
2. This repo's .claude/ folder
3. Other projects under C:\development\ (alphabetical for determinism)

Companion Command Detection

When scanning a project, a .md file in .claude/commands/ is a companion to skill X if its body matches (case-insensitive):

Invoke the `[anything containing X]` skill

or uses X:* namespace notation.

Sub-Skill Detection

A SKILL.md whose body is primarily a delegation (Invoke the \X` skill and execute...) is a sub-skill of X. It belongs in X's sub-skills/` bundle directory.

---

Skill Metadata Standards

All SKILL.md files support these optional frontmatter fields beyond the required name and description:

• status: draft | active (default) | deprecated

  • draft — skill is in development; /install-skill skips unless forced
  • active — production-ready; normal install/sync behaviour
  • deprecated — superseded or obsolete; /install-skill skips and warns; /find-skills classifies as Deprecated rather than New/Changed

• version: semver string (e.g. 1.2.0) or ISO date (e.g. 2026-04-05) — displayed by /update-skill so the user sees the version delta before accepting an update

• requires: YAML list of skill names this skill depends on (e.g. [project-manager, base]) — /install-skill reads this and auto-installs all listed dependencies before installing the target skill

---

Operation: /find-skills [path]

Discover skills across the workstation that are new or changed relative to the archive. Read-only — makes no changes.

Steps

1. Determine scan scope:

   • If path argument given: scan only that directory
   • If no argument: prompt the user to confirm before scanning:
     • Question: "Scan all source locations for new or changed skills?"
     • Options: "Yes, scan everything" | "Cancel"
     • Proceed only on confirm

2. For each discovered SKILL.md:

   • Read frontmatter — if installed-from: llm_skills is present, skip (installed copy)
   • If status: deprecated is present, classify as Deprecated regardless of diff — do not treat as New or Changed
   • Compare content against archive counterpart using diff (ignore trailing whitespace)
   • Classify: New | Changed | Unchanged | Orphan | Deprecated
   • If version: is present and archive has a different version, note the version delta in the report

3. For each New or Changed skill: also search the same project's .claude/commands/ for companion commands using the detection heuristic.

4. Output a structured report:

   ┌─ Skills Manager: Find Report ──────────────────┐
   │ New (N):        skill-a, skill-b               │
   │ Changed (N):    skill-c  (v1.0 → v1.2)         │
   │ Deprecated (N): skill-d                        │
   │ Unchanged:      N skills skipped               │
   │ Orphans:        N archive-only skills          │
   └────────────────────────────────────────────────┘
   

   For each New/Changed skill, list any detected companion commands beneath it.

---

Operation: /sync-skill [name]

Archive a specific skill (or all discovered skills) including its full bundle.

With a skill name

1. Search all canonical source locations for a skill named name (match by leaf directory name)
2. Skip any match where installed-from: llm_skills is in frontmatter
3. Apply conflict resolution if multiple sources found — take highest-priority source
4. Detect companion commands and sub-skills in the same project (see heuristics above)
5. Write to archive:
   • SKILL.md → claude/skills/<name>/SKILL.md
   • Companion commands → claude/skills/<name>/commands/<cmd>.md
   • Sub-skills → claude/skills/<name>/sub-skills/<sub>/SKILL.md
6. Generate a visual diagram for the skill:
   • Invoke the visual-explainer:generate-web-diagram skill
   • Instruct it to diagram: the skill's purpose (from description frontmatter), its key operations or phases (from section headings), significant inputs/outputs, and any decision points or branching paths
   • Save the resulting HTML to claude/skills/<name>/diagram.html
   • Add or update a ## Diagram section at the bottom of claude/skills/<name>/SKILL.md (before any final horizontal rule) containing exactly: [View diagram](diagram.html)
7. Update README: add row if new, update description if description frontmatter changed; include diagram link (see README Update Rules)
8. Report what was written

Without a skill name

1. Run the /find-skills scan internally (no user prompt for the scan itself) to enumerate scope
2. Ask the user to confirm:
   • Question: "Sync all N new/changed skills? (list skill names)"
   • Options: "Yes, sync all" | "Cancel"
3. Proceed only on confirm; process each New and Changed skill in turn

README Update Rules (sync)

• New skill: extract description from frontmatter, classify subsection (see heuristics), insert alphabetically within subsection, mark ✓ in correct toolset column, increment counts
• Changed skill: update Description cell only if frontmatter description changed; do not move the row
• Bundled commands: document on the skill's row as — ships with /cmd1, /cmd2; do NOT add standalone rows for them in the Commands table
• Sub-skills in bundles: do NOT list as standalone rows; implied by parent skill row
• Diagram link: every skill row that has a diagram.html in its archive directory must include a diagram link in the Skill column cell, formatted as: [`<name>`](claude/skills/<name>/) [(diagram)](claude/skills/<name>/diagram.html)
• Never delete from archive; never remove rows

---

Operation: /install-skill [name] [target-dir]

Deploy a skill bundle from the archive into a project.

Steps

1. Determine scope:

   • If name given: install that skill only
   • If no name: enumerate all skills in claude/skills/ that have not already been installed in target-dir; prompt the user to confirm:
     • Question: "Install all N available skills into <target>? This will write N files."
     • Options: "Yes, install all" | "Cancel"
     • Proceed only on confirm

2. For a named skill: locate claude/skills/<name>/ in the archive; if not found ask the user:

   • Question: "Skill <name> not found in archive. Check the name and try again."
   • Options: "OK" | "List available skills"

3. If target-dir not provided, ask the user:

   • Question: "Where should the skill(s) be installed?"
   • Options: "Current project directory" | "Other path (specify below)"

4. Resolve dependencies: read requires: list from each skill's frontmatter; recursively locate those skill bundles in the archive and add them to the install scope. Deduplicate — never install the same skill twice.

5. Inventory what will be installed (target skill + all resolved dependencies):

   • SKILL.md (1 file per skill)
   • Sub-skills from sub-skills/ (N files)
   • Commands from commands/ (N files)

6. If name was given, prompt the user to confirm:

   • Question: "Install <name> (+ N dependencies) into <target>? This will write N files."
   • Options: "Yes, install" | "Cancel"

7. On confirm:

   • Write <target>/.claude/skills/<name>/SKILL.md; inject installed-from: llm_skills into frontmatter (add after existing fields)
   • Write each sub-skill to <target>/.claude/skills/<sub>/SKILL.md with same marker
   • Write each command to <target>/.claude/commands/<cmd>.md (no marker — commands are not skills)

8. Report all files written

---

Operation: /update-skill [name]

Update installed skills in the current project when the archive has newer versions.

Steps

1. Scan current project's .claude/skills/ for files with installed-from: llm_skills in frontmatter

2. If name given: filter to that skill only

3. For each installed skill found:

   • Compare content against archive version
   • If identical: skip (no update needed)
   • If archive is newer: record as outdated

4. Confirm scope with the user before touching anything:

   • If updating all: "Update all N outdated skills? (list skill names, include version delta where available, e.g. skill-x v1.0→v1.2)"
   • If updating one by name: "Update <name> to the archive version? (v1.0 → v1.2)"
   • Options: "Yes, update" | "Cancel"

5. On confirm:

   • Overwrite each SKILL.md, preserving the installed-from: llm_skills marker in frontmatter
   • Also check commands/ in the archive bundle — write any new or changed commands to <project>/.claude/commands/
   • Also check sub-skills/ — update any installed sub-skills that are outdated
   • Regenerate the diagram: invoke visual-explainer:generate-web-diagram for each updated skill and overwrite claude/skills/<name>/diagram.html; update the ## Diagram section in SKILL.md if the path changed

6. Report updated files

---

Operation: /import-skill [name]

Import project-level changes to a skill back into the archive and the global user profile.

Steps

1. Determine scope:

   • If name given: import that skill only
   • If no name: scan current project's .claude/skills/ for all skills that do NOT have installed-from: llm_skills and differ from the archive; prompt the user to confirm:
     • Question: "Import all N project-developed skills to archive and user profile? (list names)"
     • Options: "Yes, import all" | "Cancel"
     • Proceed only on confirm

2. Locate <name>/SKILL.md in the current project's .claude/skills/

3. Check frontmatter: if installed-from: llm_skills is present, ask the user to confirm rejection:

   • Question: "<name> was installed from the archive — it is not a project-developed skill. Import would just overwrite the archive with its own content. Proceed anyway?"
   • Options: "Cancel" | "Yes, force import"
   • Default is Cancel; only proceed if user explicitly chooses force

4. Diff current project version against archive version

5. Show summary: lines added, lines removed, key description change (if any)

6. Prompt the user to confirm:

   • Question: "Import <name> changes to archive and user profile? (N lines added, N removed)"
   • Options: "Yes, import both" | "Archive only" | "Cancel"

7. On confirm:

   • Archive: overwrite claude/skills/<name>/SKILL.md
   • User profile: overwrite ~/.claude/skills/<name>/SKILL.md (create dir if needed)
   • Companion commands: scan project .claude/commands/ for new companion commands not yet in the archive bundle; for each new one, ask the user: "Bundle /<cmd> with <name>?" → "Yes" | "Skip"
   • README: update Description cell if frontmatter description changed

8. Report what was imported

---

Operation: /backfill-diagrams [name]

Generate diagram.html for archived skills that don't have one, retroactively applying the diagram standard.

Steps

1. Determine scope:

   • If name given: check only that skill
   • If no name: scan claude/skills/*/ and collect every skill directory that lacks diagram.html

2. Prompt the user to confirm:

   • Question: "Generate diagrams for N skills that are missing one? (list skill names)"
   • Options: "Yes, backfill all" | "Cancel"
   • Proceed only on confirm

3. For each skill in scope:

   • Invoke visual-explainer:generate-web-diagram using the skill's purpose, operations, and key decision points as input
   • Save HTML to claude/skills/<name>/diagram.html
   • Add or update ## Diagram section in claude/skills/<name>/SKILL.md with [View diagram](diagram.html)

4. Update README: for each skill now having a diagram, add the diagram link to its Skill column cell (see README Update Rules)

5. Report how many diagrams were generated

---

Operation: /search-skill [query]

Search the archive for skills matching a keyword or phrase. Read-only — no changes made.

Steps

1. If no query provided, ask the user:

   • Question: "What are you looking for?"
   • Options: "Enter search query" (free text via Other)

2. Search the following fields across all claude/skills/*/SKILL.md files (case-insensitive):

   • name: frontmatter field — highest priority
   • description: frontmatter field — high priority
   • ## section headings in the body — medium priority
   • Body text — lower priority

3. Rank results: exact name match > description keyword hit > heading hit > body hit. Break ties alphabetically.

4. Output ranked results:

   ┌─ Skills Manager: Search Results for "auth" ────┐
   │ 1. security          (Security & Credentials)   │
   │    "OWASP security patterns, secrets..."        │
   │    [diagram](claude/skills/security/diagram.html)│
   │ 2. credentials       (Security & Credentials)   │
   │    "Centralized API key management..."          │
   │ 3. supabase          (Databases & Storage)      │
   │    "...Auth, Storage, real-time..."             │
   └────────────────────────────────────────────────┘
   

   Include diagram link on result lines where diagram.html exists.

---

Operation: /audit-skills

Run a comprehensive read-only health check across the entire archive. No changes made, no confirmation needed.

Checks

1. Missing diagrams: skills in claude/skills/*/ that lack diagram.html
2. README drift — broken links: rows in the README Skills table whose link target directory does not exist in claude/skills/
3. README drift — missing rows: skill directories in claude/skills/*/ that have no corresponding row in the README Skills table (match by link text)
4. Malformed frontmatter: SKILL.md files missing required name: or description: fields
5. Deprecated installs: scan all projects under C:\development\ for installed skills (installed-from: llm_skills) where the archive copy has status: deprecated — list project and skill name
6. Codex/Gemini parity gaps: README rows with ✓ only in the Claude column — grouped by subsection, sorted by description length as a proxy for skill complexity (longer = more porting value)

Output format

┌─ Skills Manager: Audit Report ─────────────────────────────────┐
│ Missing diagrams (N):   skill-a, skill-b, ...                  │
│ README broken links (N): row-name → path/does/not/exist        │
│ README missing rows (N): skill-dir has no README entry         │
│ Bad frontmatter (N):    skill-c (missing description)          │
│ Deprecated installs (N): project-x uses deprecated skill-d     │
│ Parity gaps — Claude only (N skills, top candidates):          │
│   - security (Security & Credentials)                          │
│   - typescript (Languages & Runtimes)                          │
│   ...                                                           │
└────────────────────────────────────────────────────────────────┘

---

Operation: /push-skill [name]

Push a skill bundle from the archive to the global user profile (~/.claude/) so it is available across all projects.

Steps

1. Determine scope:

   • If name given: push that skill bundle only
   • If no name: enumerate all skills in claude/skills/; prompt the user to confirm:
     • Question: "Push all N archive skills to the global profile (~/.claude/)?"
     • Options: "Yes, push all" | "Cancel"
     • Proceed only on confirm

2. For each skill to push:

   • Check if status: draft — skip with a warning; drafts are not pushed to the global profile
   • Check if ~/.claude/skills/<name>/SKILL.md already exists and differs — ask the user:
     • Question: "~/.claude/skills/ exists and differs. Overwrite with archive version?"
     • Options: "Yes, overwrite" | "Skip this skill"

3. On confirm:

   • Write SKILL.md to ~/.claude/skills/<name>/SKILL.md (do NOT add installed-from marker — the global profile is a source, not a deployment target)
   • Write each sub-skill from sub-skills/ to ~/.claude/skills/<sub>/SKILL.md
   • Write each companion command from commands/ to ~/.claude/commands/<cmd>.md

4. Report all files written

---

Source Discovery Table

Source	Path Pattern	Toolset	Archive Destination
Global profile	~/.claude/skills/*/SKILL.md	Claude	claude/skills/<name>/SKILL.md
Global profile	~/.claude/agents/*.md	Claude	claude/agents/<name>.md
Global profile	~/.claude/commands/*.md	Claude	claude/commands/<name>.md
This repo	.claude/skills/*/SKILL.md	Claude	claude/skills/<name>/SKILL.md
This repo	.claude/commands/*.md	Claude	claude/commands/<name>.md
Other projects	.claude/skills/*/SKILL.md	Claude	claude/skills/<name>/SKILL.md
Other projects	.claude/agents/*.md	Claude	claude/agents/<name>.md
Other projects	.claude/commands/*.md	Claude	claude/commands/<name>.md
Other projects	skills/*/SKILL.md	Claude	claude/skills/<name>/SKILL.md
Other projects	codex/skills/*/SKILL.md	Codex	codex/skills/<name>/SKILL.md
Other projects	gemini/skills/*/SKILL.md	Gemini	gemini/skills/<name>/SKILL.md

Nested paths: The skill <name> is always the leaf directory containing SKILL.md. Intermediate path segments (e.g. homeradar/ in skills/homeradar/feature-start/SKILL.md) are discarded.

Installed-copy filter: Apply before any classification. Skip any file with installed-from: llm_skills in frontmatter.

llm_skills itself: Only scan .claude/ — never recurse into claude/ (the archive destination). Scanning the archive as a source would create circular copies.

Directories to skip: nul, emby-opbta.p12, fullchain.pem, privkey.pem, recovery_app, and any loose files (not directories) at the C:\development\ root.

---

README Classification Heuristics

Infer subsection from keywords in the skill's description field:

Keywords	Subsection
git, branch, commit, PR, merge, workflow, session, team, coordination, repo	Foundations & Workflow
TypeScript, Python, Node.js, runtime, language	Languages & Runtimes
React, Vue, Next.js, frontend framework, PWA, Flutter	Frontend Frameworks
Android, iOS, mobile native, Kotlin, Swift	Mobile (Native)
UI, UX, design, Tailwind, accessibility, visual, diagram, explainer	UI & Design
database, SQL, Supabase, Firebase, DynamoDB, Aurora, Cosmos, D1, schema	Databases & Storage
code review, testing, quality, lint, coverage, Playwright, duplication	Code Quality
security, credentials, auth, OWASP, secrets, vulnerability	Security & Credentials
AI, LLM, agent, prompt, model, Anthropic, OpenAI, Gemini	AI & LLM
commerce, Shopify, WooCommerce, Medusa, Stripe, payment	Commerce & Payments
Klaviyo, Reddit, Teams, PostHog, third-party, integration	Third-Party Integrations
SEO, sitemap, robots, web presence, content, GEO	SEO & Web Presence
tooling, DevOps, CLI, deploy, Vercel, render, CI/CD	Tooling & DevOps
OSINT, research, intelligence, GEOINT, worldview, forensics	Research & OSINT

If no keywords match: Uncategorized. Multiple matches: most keyword hits wins; break ties with more specific subsection.

---

Archive Invariants — Never Violate

1. Never delete from archive. Even if a source file no longer exists, the archive copy stays. Deletions are manual curatorial decisions.
2. Flow is always source → archive — except /import-skill, which explicitly reverses this.
3. The archive README is the authoritative index. Every archived skill must have a row; every row must point to a real file.
4. Toolset is determined by directory, not content.
5. Installed copies are not project-developed skills. The installed-from: llm_skills marker gates this distinction.