Intel / article

Linting 216 public Claude Code skills – 69% won't reliably trigger

Source: skillcrossroads.com
Author: sgharlow
Date: Published Sep 17, 2026

Context

A developer built a linter for Claude Code artifacts, skills, subagents, slash commands, configs, and plugins, then ran it against public repositories to see what people actually ship (published audit). Across 215 public skills scored for triggering, 69 percent had a description unlikely to reliably fire, 40 percent outright unlikely and 28 percent borderline, which the author identifies as the most common real-world failure mode, hidden in a single line of frontmatter. On the side, 57 percent of 87 sampled subagents declared no tools list at all, which reads as a safe default but isn't: a subagent with no declared tools inherits the caller's entire toolbox, Bash included, the opposite of least-privilege. Permission prompts still gate execution, as the subagent audit notes. Combined with 16 subagents that grant bare Bash access and 8 that grant a wildcard, 85 percent of the sample fell short of least-privilege tool scoping. Only 1 of 216 skills and none of the 87 subagents passed every check cleanly.

Key takeaways · AI-distilled

57% of the 87 sampled subagents (50) declare no `tools` list at all, which does not them, it makes them inherit the caller's entire toolbox including Bash, the opposite of least privilege.
Adding subagents that grant bare Bash (16) or a wildcard (8) to the no-list group brings 85% of the 87-subagent sample to broader-than-needed tool access.
Of 215 public skills scored for triggering reliability, 40% are outright unlikely to fire and another 28% are borderline, so 'my never fires' traces to one frontmatter description line in most cases.
Only 1 of 216 skills and 0 of 87 subagents passed every check in the linter cleanly; 83% of subagents (72 of 87) also omit 'use when...' invocation cues from their description.
Secrets handling was the one area sampled artifacts got mostly right: only 6 of 216 skills tripped the scanner.

Terms in this piece · Glossary

MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
multi-agent — Using several AI agents on one problem — splitting work in parallel, checking each other, or filling different roles like planner and reviewer.
agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.
sandbox — An isolated environment where AI-generated code or agent actions run without being able to touch anything real.

Read the source skillcrossroads.com