mxhelp

name: mxHelp description: Use when the user says "/mxHelp", "/mxHelp ", or asks which mx-skill does what. Lists all installed mx*-skills grouped by category, or explains one skill in detail. Pure reader — no side effects. allowed-tools: Glob, Grep, Read

/mxHelp — mx-Skill Discoverability (AI-Steno: !=forbidden →=use ⚡=critical)

Thin renderer over existing SKILL.md files. NO content authoring, NO MCP writes, NO state changes. Source of truth = each ~/.claude/skills/mx*/SKILL.md. This skill only globs, parses frontmatter, and renders.

Modes

Mode 1: List (default, --verbose for full descriptions)

⚡ Tool budget: max 2 tool calls (1× Glob + 1× Grep). NO per-file Reads. NO Bash. Each Read on a fresh session triggers a permission prompt — multiplied by N skills this is unusable. Stay within Glob+Grep.

• Zero-result handling: If Glob returns zero matches OR all matches are filtered out by the allowlist, render exactly: No public mx*-skills installed. Run /mxSetup to install the mxLore-skills bundle. and return. Do NOT render empty category headers.

1. Glob (1 call): ~/.claude/skills/mx*/SKILL.md — case-insensitive on Windows NTFS. Confirms which skills exist on disk.
2. Grep (1 call) — extract all frontmatter at once:
   • pattern: '^(name|description):'
   • path: '~/.claude/skills'
   • glob: 'mx*/SKILL.md'
   • output_mode: 'content'
   • -n: true (line numbers)
   • Result: every name: and description: line for ALL mx-skills in a single tool call. Format: <filepath>:<lineno>:<content>
   • Malformed frontmatter handling: if a SKILL.md is missing name: or description: → render as <dirname> — (no frontmatter) and continue. NEVER abort the entire list. NEVER hallucinate field values.
3. Filter against allowlist: parse the grep output, group name/description pairs by file path. For each, compare the name value against the Public Skill Allowlist (case-insensitive). Skills NOT in allowlist → silently skip (private/customer-only).
4. Render compact bulleted list grouped by category from the allowlist:

   ## Core workflow
   - **mxPlan** — <first sentence of description>
   - **mxSpec** — ...
   

5. Default truncation: stop at first '.', '!', '?' OR at 120 UTF-8 codepoints (NEVER split a multi-byte sequence). Whichever comes first.
6. With --verbose: full description, no truncation.
7. Footer: Tip: /mxHelp <name> for details on one skill.

Public Skill Allowlist

⚡ mxHelp ONLY renders skills from the public mxLore-skills GitHub bundle. Internal/customer-specific skills (mxerptrigger, mxTicketSystemHannes, mxDelphiAnsi2UTF8, etc.) are FILTERED OUT — they are not part of the public distribution and must not appear in mxHelp output.

Total public skills: 12. Any glob match whose frontmatter name is NOT in this allowlist (case-insensitive compare) → SILENTLY SKIPPED. Do NOT render under "Other". Do NOT emit drift warnings for non-allowlisted skills — they are intentionally private.

Drift warning ONLY fires when an allowlisted skill is missing from disk: Note: public allowlist references N missing skill(s): <names> — this surfaces real distribution gaps.

Mode 2: Detail (/mxHelp <skillname>)

1. Sanitize <skillname>: REJECT input containing /, \, .., or any absolute-path marker. Allowed pattern: [A-Za-z0-9_-]+ only. On reject → respond "Invalid skill name: only alphanumeric + underscore + hyphen allowed" and stop.
2. Normalize: strip trailing .md, strip leading /. If no mx prefix → prepend mx and retry exact match before fuzzy.
3. Allowlist check: if the normalized name is NOT in the Public Skill Allowlist → respond "Skill <name> is not part of the public mxLore-skills bundle" and stop. Internal skills must not be discoverable via mxHelp.
4. Resolve path: ~/.claude/skills/<skillname>/SKILL.md (case-insensitive match)
5. ∅exact match → glob ~/.claude/skills/mx* → Fuzzy match algorithm (specified):
   1. Case-insensitive substring match against allowlist names → if 1+ matches, list up to 3 sorted by name length (shortest first)
   2. If 0 substring matches → Levenshtein distance ≤3 against allowlist → list up to 3 sorted by distance ascending
   3. If still 0 matches → respond "No public mx-skill matches <query>. Try /mxHelp for the full list."
   4. Tie-break: alphabetical
6. Read full file
7. Render structured sections:
   • Purpose — first paragraph after the H1 heading
   • Modes / Arguments — extract from any ## Modes table or ## Mode N: headings
   • Examples — first fenced code block after an "Example" or "Usage" heading; omit section entirely if none found
8. Footer: Source: ~/.claude/skills/<name>/SKILL.md

⚡ Detail mode is read-only. NO side effects. NO state writes. NO MCP calls.

Rules

• ⚡ Tool budget — List mode: max 2 calls (1 Glob + 1 Grep). Detail mode: max 2 calls (1 Glob + 1 Read). Exceeding the budget = bug. NEVER use Bash. NEVER use per-file Read in List mode.
• ⚡ Output language: match the user's prompt language. This SKILL.md is in English for sync hygiene (public mxLore-skills GitHub repo requires English), but the rendered output (category headers, descriptions, footer tip, error messages) MUST be translated to match the user's language. If the user wrote in German → render German. Italian → Italian. English → English. Default English when the language is ambiguous.
• !duplicate skill content — only render what already exists in SKILL.md
• !invent skills — if an allowlisted skill does not exist on disk, omit from render AND append a one-line footer warning (in the user's language): Note: public allowlist references N missing skill(s): <names> (surfaces drift without breaking reads)
• Include mxHelp itself in the list under "Discoverability"