llama-cpp

Runs LLM inference on CPU, Apple Silicon, and consumer GPUs without NVIDIA hardware. Use for edge deployment, M1/M2/M3 Macs, AMD/Intel GPUs, or when CUDA is unavailable. Supports GGUF quantization (1.5-8 bit) for reduced memory and 4-10× speedup vs PyTorch on CPU.

Quality

83%

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

Quality

Content

77%

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

The body is highly actionable with executable code and clear quick-start sequencing, and it correctly delegates to real reference files. Its main weakness is conciseness/progressive-disclosure balance: benchmark tables and a full supported-models catalog are inlined where the references could absorb them.

Suggestions

Move the CPU and GPU performance benchmark tables into references/optimization.md and keep only a one-line pointer in the body.

Relocate the 'Supported models' catalog to references/quantization.md (or a dedicated models reference), leaving a brief summary plus the HuggingFace link inline.

Trim the 'Use TensorRT-LLM/vLLM instead when' comparison to a single 'Use this for non-NVIDIA hardware' note, since the description already establishes the niche.

Dimension	Reasoning	Score
Conciseness	Mostly efficient and code-dense, but the ~248-line body retains some unnecessary inline material (benchmark tables, a full supported-models catalog, a 'Use TensorRT/vLLM instead when' comparison) that competes with the context budget.	2 / 3
Actionability	Executable, copy-paste-ready bash commands with specific flags throughout (e.g. 'make LLAMA_METAL=1', '-ngl 32', '--grammar-file grammars/json.gbnf') match the fully-executable anchor.	3 / 3
Workflow Clarity	Installation, download, inference, and server steps are clearly sequenced with inline comments; the operations are non-destructive so the missing validation checkpoints do not warrant a cap.	3 / 3
Progressive Disclosure	Real one-level-deep references are present and well-signaled (references/quantization.md, server.md, optimization.md), but substantial inline content (benchmark tables, full model catalog) that could live in those references keeps it from a 3.	2 / 3
	Total	10 / 12 Passed

Description

90%

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 strong, third-person description with explicit 'what' and 'when' guidance and natural trigger terms that clearly carve out a non-NVIDIA inference niche. The only slight gap is specificity, since the named concrete actions are limited to inference and quantization.

Dimension	Reasoning	Score
Specificity	It names concrete actions ('Runs LLM inference', 'Supports GGUF quantization (1.5-8 bit) for reduced memory') but the action set is narrow rather than the comprehensive multi-action list required for a 3.	2 / 3
Completeness	Explicitly answers both what ('Runs LLM inference... Supports GGUF quantization') and when ('Use for edge deployment, M1/M2/M3 Macs... or when CUDA is unavailable'), satisfying the explicit-trigger requirement.	3 / 3
Trigger Term Quality	Strong natural coverage of terms users would actually say — 'CPU', 'Apple Silicon', 'M1/M2/M3 Macs', 'AMD/Intel GPUs', 'edge deployment', 'CUDA is unavailable' — matching the good-coverage anchor.	3 / 3
Distinctiveness Conflict Risk	Carves a clear niche (non-NVIDIA / CUDA-unavailable inference) with a distinguishing trigger unlikely to fire for NVIDIA-focused alternatives.	3 / 3
	Total	11 / 12 Passed

Validation

93%

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 — 15 / 16 Passed

Validation for skill structure

Criteria	Description	Result
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning

	Total	15 / 16 Passed

Repository: Orchestra-Research/AI-Research-SKILLs
Path: 12-inference-serving/llama-cpp/SKILL.md
Commit: 773a529

Reviewed: about 12 hours 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.