llamaguard

Meta's 7-8B specialized moderation model for LLM input/output filtering. 6 safety categories - violence/hate, sexual content, weapons, substances, self-harm, criminal planning. 94-95% accuracy. Deploy with vLLM, HuggingFace, Sagemaker. Integrates with NeMo Guardrails.

Quality

61%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Low

Low-risk findings worth noting

Fix and improve this skill with Tessl

tessl review fix ./backend/cli/skills/llm-tools/llamaguard/SKILL.md

Quality

Content

62%Weight 40%Scale 1-5

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

Highly actionable, executable content with strong code coverage across deployment paths, but it over-explains basics and inlines a lot of material that the reference links (which are missing files) are supposed to offload. Batch/moderation workflows also lack validation checkpoints, capping workflow clarity.

Suggestions

Create the referenced files (references/custom-categories.md, references/benchmarks.md, references/deployment.md) or remove the dead links so navigation is not broken.

Add validation in the moderation helpers — guard against 'unsafe' output with no category line, and validate each result in the batch moderation example.

Trim boilerplate explanations (install commands, memory math, what quantization does) that Claude already knows; keep only model-specific gotchas.

Dimension	Reasoning	Score
Conciseness	Mostly efficient code-first content, but it explains some things Claude already knows (installing transformers, what tensor_parallel_size does, that 8-bit halves memory) and pads with repeated model-init snippets across workflows.	3 / 5
Actionability	Provides fully executable, copy-paste-ready code for the common cases (HF transformers, vLLM, FastAPI endpoint, NeMo Guardrails) plus concrete commands and curl examples; specific outputs are shown inline.	5 / 5
Workflow Clarity	Workflows are clearly sequenced into numbered stages, but input/output moderation and batch processing lack validation checkpoints (no handling of malformed 'unsafe' output missing a category, no empty-result guards), and the batch moderation example performs no validation per result.	3 / 5
Progressive Disclosure	The body inlines all substantive content (5 full workflows, API reference, hardware specs) while pointing to references/custom-categories.md, references/benchmarks.md, and references/deployment.md — but those reference files do not exist, so the navigation is broken and inlining is heavy.	2 / 5
	Total	13 / 20 Passed

Description

60%Weight 40%Scale 1-5

Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.

A reasonably specific and distinctive description that names the model, categories, accuracy, and deployment options, but it omits any explicit 'Use when...' trigger guidance, which caps completeness. Adding a trigger clause and more natural user phrasings would lift it.

Suggestions

Add an explicit 'Use when...' clause, e.g. 'Use when moderating LLM inputs/outputs for safety, classifying harmful content, or filtering prompts/responses.'

Surface natural user phrasings users would actually say ('is this message safe', 'block harmful prompts', 'moderate responses').

Keep third-person voice (already compliant) but trim the accuracy/deployment detail into the body to keep the description as a trigger-focused overview.

Dimension	Reasoning	Score
Specificity	Lists several concrete capabilities (input/output filtering, 6 named safety categories, deployment with vLLM/HuggingFace/Sagemaker, NeMo Guardrails integration), though some are deployment targets rather than actions the skill performs.	4 / 5
Completeness	The 'what' is clear (moderation model with 6 categories, accuracy, deployment options) but there is no 'Use when...' clause or equivalent trigger guidance, which caps completeness at 3 per the guidelines.	3 / 5
Trigger Term Quality	Includes strong natural terms like 'content moderation', 'input/output filtering', 'safety categories', and 'LlamaGuard', but misses common user phrasings like 'moderate this text', 'is this safe', or 'block harmful content'.	4 / 5
Distinctiveness Conflict Risk	It carves a clear niche (Meta LlamaGuard moderation specifically) with minimal overlap risk against generic safety skills, though 'content moderation' broadly could overlap with OpenAI/Perspective alternatives.	4 / 5
	Total	15 / 20 Passed

Validation

81%

Warnings & errors only

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation — 13 / 16 Passed

Validation for skill structure

Criteria	Description	Result
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning
relative_links	Relative link issues: 3 missing	Warning
referenced_paths_exist	Referenced path issues: 6 missing	Warning

	Total	13 / 16 Passed

Repository: synthetic-sciences/openscience
Path: backend/cli/skills/llm-tools/llamaguard/SKILL.md
Commit: fcd2e11

Reviewed: 2 days ago

Table of Contents

Discovery Implementation Validation

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.