NLPM Audit: 4 bugs and 3 security improvements found (score: 83/100)

Automated audit: This issue was generated by NLPM, a natural language programming linter, running via claude-code-action. Please evaluate the findings on their merits.

About This Audit

Agent Skills for Context Engineering is a high-quality collection — it scored 83/100 on NLPM's natural-language programming rubric, with 12 of the 14 core skills earning 90 or above. This audit surfaces the few actionable issues worth fixing; everything else is strong work.

NLPM is a 100-point quality scoring tool for Claude Code NL artifacts (skills, agents, hooks). It applies deterministic penalties against a rulebook and flags bugs that affect discoverability or correctness. An automated pipeline then opens PRs for verified bugs.

Bugs Found (PR-worthy)

These are structural bugs that break functionality, not style nits.

Priority 1 — Missing YAML frontmatter in llm-as-judge-skills agents (4 files)

Your CONTRIBUTING.md requires YAML frontmatter with name and description fields in every skill/agent file. All four files in examples/llm-as-judge-skills/agents/ are missing this block entirely, making them invisible to Claude Code's skill loader.

Fix PR: #80

Security Improvements Found (Medium/Low only)

Priority 2 — GitHub token embedded in git clone URL (Medium)

skills/hosted-agents/scripts/sandbox_manager.py line 186 constructs a git clone URL with the token directly in the string:

f"git clone https://x-access-token:{token}@github.com/{repo_url} /workspace"

Even though this file is documented as pseudocode, this pattern teaches an unsafe practice: the token appears in ps aux, shell history, and any command logging infrastructure.

Fix PR: #81

Priority 3 — Unpinned production dependency ranges (Low)

examples/llm-as-judge-skills/package.json uses caret (^) ranges for all 5 production deps, allowing automatic minor version upgrades and raising supply-chain risk.

Fix PR: #82

Priority 4 — Unsanitized user path input in install.sh (Low)

examples/digital-brain-skill/scripts/install.sh line 37 passes the user-supplied custom path directly to mkdir -p and cp -r without validation, allowing .. traversal.

Fix PR: #83

Quality Notes (informational, no PRs)

The following are quality observations only — no action required:

• examples/digital-brain-skill/agents/AGENTS.md: No declared model, no agent I/O examples, no output format spec. Still scored 70/100.
• examples/interleaved-thinking/generated_skills/comprehensive-research-agent/SKILL.md: Formatting bug (*Silent should be **Silent in several bullet items) and missing version field in frontmatter.
• examples/llm-as-judge-skills/agents/orchestrator-agent/orchestrator-agent.md: Lists writer and analyst agents in Available Agents, but no agent directories exist for them. Minor documentation inconsistency.
• skills/hosted-agents/SKILL.md: Strong guidelines and gotchas, but no ## Examples section with concrete I/O pairs.
• skills/latent-briefing/SKILL.md: Only one example scenario; a second would improve clarity.

Thank you for a well-organized and useful collection. The core skills are exemplary — several scored 95/100. These fixes should get the overall score above 90.