help

Passive on-demand index of installed user-invoke skills, grouped by category. Bundled with super-bootstrap. Invoke as `/super-bootstrap:help` (namespaced to avoid colliding with Claude Code's built-in

SKILL.md

GitHub

Help — User-Invoke Skill Discovery Surface

Render an on-demand menu of slash commands the user can invoke in this project. Cold-by-nature user-invoke skills (slash commands the user must remember) need a discovery surface; passive /super-bootstrap:help is the right shape.

Extraction is mechanical (structured JSON + YAML frontmatter), so a bundled script does it in one tool call with zero model tokens; the one judgment in the pipeline — which discovered skills a user would actually invoke — stays with the gateway.

When to Use

  • User forgot which slash commands are available
  • User just installed a new plugin and wants to see what it added
  • User wants to filter by category (/super-bootstrap:help git, /super-bootstrap:help docs)

Execution (gateway-inline)

When the user invokes /super-bootstrap:help [category]:

  1. Extract — run the bundled script, passing the project root (the session's working directory — pwd): python3 "${CLAUDE_PLUGIN_ROOT}/skills/help/assets/render-menu.py" "$(pwd)" (The script body is opaque to the permission engine — the grant to carve is the invocation itself, e.g. Bash(python3 *).) It emits one candidate per line (category<TAB>command<TAB>description) from the installed-plugin registry, enabledPlugins, each enabled plugin's skills, and project .claude/skills. It deliberately over-reports — no filter lives in the script.
  2. Filter to user-invoke — drop rows a user would never type: delegation-only skills (dispatched by other skills or the model, e.g. process/reference skills another skill invokes), hook-fired skills, and reference-material bundles. Keep anything a user plausibly types as a slash command. This is the judgment step; do it from the emitted descriptions — no file reads.
  3. Render — group surviving rows under their category headers ([meta] [pipeline] [git] [docs] [dev] [utils]), one line per skill: command + one-line summary. Header: Available slash commands ({N} total):. A category argument renders only that category; unknown category → list available categories. Anomaly footer: the script's stderr carries an # anomalies: line only when a source emitted fewer rows than it declared or held — render everything after its # anomalies: prefix as a footer line beneath the menu, contents unaltered and itself prefixed incomplete: . A short source is a discovery failure, and without the footer it is indistinguishable from a plugin that ships no commands. No # anomalies: line → no footer.
  4. Fallback — python3 missing on the device: scan the same sources manually with Read/Grep (structure documented in the script header) and render the same shape.

Rules

  • No active reminders. Discovery is pull-only — the user invokes; nothing fires ambiently.
  • Namespaced invocation. Always /super-bootstrap:help — bare /help is Claude Code's built-in. Footer hints elsewhere in this plugin must use the namespaced form too.
  • Script over-reports by design. The user-invoke filter is the gateway's judgment at render time; never push it down into the script.