Skill Crossroads — bash-expert

Skill Crossroads bash-expert
rubric v1.2 · deterministic

overall

Correctness & Structure 100 pass

Triggering & Discoverability 85 1 warn

Clarity & Instructions 100 pass

Token & Context Cost 100 pass

Safety & Security 81 2 warn

Verifiability & Maintainability not yet scored

Top fixes — ranked by score impact

⚠ SAFETY-03 No destructive auto-invocation

bash-expert.md:35auto-invocable (no disable-model-invocation) → performs a destructive action ("rm -rf")

Skill appears to "rm -rf" but the model can auto-invoke it — a destructive action could fire without the user asking.

rm -rf

Fix If this action is destructive/irreversible, set `disable-model-invocation: true` so it only runs when explicitly invoked.

⚠ TRIGGER-03 Invocation cues in the description

bash-expert.md:1description will match user requests → no explicit cue ('use when …', 'triggers on …', quoted user phrases) found

Pure what-it-does prose — the model has to guess when to invoke it.

Master of defensive Bash scripting for production automation, CI/CD pipelines…

Fix State the trigger explicitly: 'Use when the user says "…", "…", or wants to …'.

⚠ SAFETY-02 allowed-tools least-privilege

bash-expert.md:1no `tools` declared → the agent inherits every tool, including Bash

Subagents without a `tools` list get unrestricted tool access.

Fix Declare a scoped `tools:` list with only what the agent needs (e.g. `tools: Read, Grep, Glob`).

Partial grade — some applicable rubric categories were not scored on this scan (e.g. LLM-assisted checks without a key) and are excluded from the overall.

Embed this badge

Always-fresh — it re-scans and updates on its own.

[![Skill Crossroads](https://skillcrossroads.com/api/badge/0xfurai/claude-code-subagents/agents/bash-expert.md.svg)](https://skillcrossroads.com/s/0xfurai/claude-code-subagents/agents/bash-expert.md)

Grade your own Claude Code skill — evidence-cited, file-and-line, free. Scan your own skill →