// the find
amElnagdy/guard-skills
Guard skills for coding agents, quality gates that catch AI-generated failure modes in code, tests, and docs
A set of review-prompt 'skills' for AI coding agents — clean-code, test, docs, WordPress, and WooCommerce guards — installed via the skills.sh CLI and invoked as a second pass after the agent produces a diff. The idea is to catch systematic LLM failure modes (silent error swallowing, hallucinated APIs, copy-paste test bloat) before they land in your repo. Aimed squarely at teams already using Claude Code, Codex, or Cursor.
- The reactive framing is correct: running a guard on a finished diff gets better results than trying to constrain generation upfront, and the README makes this distinction clearly instead of overpromising.
- Progressive-disclosure structure — SKILL.md stays small, reference files load only when the task needs them — is real prompt engineering that respects context window costs, not just a nice folder layout.
- The AI-specific layer in clean-code-guard (catch-all swallowing, hardcoded success returns, copy-from-similar bugs, hallucinated APIs) targets failure modes that no static linter catches. This is the part that actually justifies the repo's existence.
- Fully inspectable: pure Markdown, no executables, no MCP server, no network calls during use. You know exactly what instructions your agent receives, which matters when you're handing a skill root access to your codebase.
- Entirely dependent on the skills.sh ecosystem — a Vercel Labs project with no guaranteed longevity. If that CLI changes its discovery format or goes unmaintained, these become manually-copied Markdown files with no install story.
- There's no way to verify the guards actually work from the repo. No recorded examples of caught failures, no evaluation harness, no before/after diffs. You're taking the author's word that these prompts change agent behavior in measurable ways.
- Two of five guards (wp-guard, woo-guard) target a single ecosystem that's a minority use case for AI coding agents. That's 40% of the skill surface covering WordPress/WooCommerce, which looks like the maintainer's professional background driving scope rather than general demand.
- The update model is manual and has no integrity story. Skills install as a copy; to get updates you re-run a CLI command that fetches from a registry you don't control. There's no checksum, no pinned version, no diff-before-apply — you're trusting the registry hasn't changed the skill in a way you wouldn't endorse.