What it is
Seven custom subagents for Claude Code, as markdown files you drop into
.claude/agents/: code-refactorer, content-writer, frontend-designer, prd-writer,
project-task-planner, security-auditor, and vibe-coding-coach. Each is a scoped
system prompt the main agent can delegate to.
How it was vetted
Read every agent definition at commit f7df2c3 (the repo’s only commit,
2025-07-25) and ran the injection audit.
- Flagged for over-broad tool grants. content-writer, frontend-designer, and
vibe-coding-coach carry no
tools:line, so a delegated run inherits all tools, Bash and web included. security-auditor is granted Write and Edit but no Read. project-task-planner’s own prompt says it may not action tasks, yet it holds Bash, Write, and Edit. - Stale packaging. One commit from 2025, deprecated tool names (
MultiEdit,LS), and two agents (frontend-designer, content-writer) whose in-prompt examples use names that no longer match the file. - Description bloat. Each delegation description embeds full example transcripts, about 375 always-on tokens per agent, where a one-line trigger would do.
- No malware. No exfiltration, no phone-home, no credential reads, no injection. The one external link is a referral to the author’s own site in project-task-planner. The prompts themselves are substantive: a real OWASP-style checklist in security-auditor, full section schemas in the writers.
- The fix. Copy only the agents you have read, and pin each one’s
tools:to least privilege.
