sglang

Fast structured generation and serving for LLMs with RadixAttention prefix caching. Use for JSON/regex outputs, constrained decoding, agentic workflows with tool calls, or when you need 5× faster inference than vLLM with prefix sharing. Powers 300,000+ GPUs at xAI, AMD, NVIDIA, and LinkedIn.

Quality

82%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Critical

Do not install without reviewing

Quality

Content

65%

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

A thorough, highly actionable SGLang guide with excellent executable examples and well-signaled one-level references, but it is overlong and somewhat redundant — benchmark tables, model/hardware lists, and the RadixAttention explainer duplicate content that belongs in the existing reference files. Adding a brief server-health validation step and trimming inline detail would lift the weaker dimensions.

Suggestions

Trim the body to a lean overview: move the full benchmark tables, supported-models/hardware lists, and the RadixAttention 'how it works' explainer into the existing reference files (radix-attention.md, deployment.md) and link to them instead of inlining.

Reduce the structured-generation section to one canonical pattern inline and defer the additional regex/grammar/schema variants to structured-generation.md, since that reference already exists.

Add an explicit validation checkpoint to the quick-start workflow — e.g. curl the /health endpoint to confirm the server is ready before running inference — so the launch→infer sequence has a feedback step.

Dimension	Reasoning	Score
Conciseness	Mostly actionable code but ~430 lines with redundancy and explanation Claude partly already knows (the full RadixAttention 'how it works' walkthrough, benchmark tables, supported-models/hardware lists), so it could be tightened rather than being lean throughout.	2 / 3
Actionability	Provides extensive copy-paste-ready, executable guidance throughout — install commands, server launch flags, and complete @sgl.function code for inference, JSON/regex/grammar, agents, multi-turn, multi-modal, batching, and the OpenAI API.	3 / 3
Workflow Clarity	A clear install→launch→infer sequence exists in Quick start, but the body is organized as a feature catalog rather than a sequenced workflow and includes no validation checkpoints or feedback loops.	2 / 3
Progressive Disclosure	References are real, clearly signaled, and one level deep (structured-generation.md, radix-attention.md, deployment.md), but the body is overlong with inline content — full benchmark tables, model/hardware lists, and a complete RadixAttention explainer — that overlaps the reference files and should be split out.	2 / 3
	Total	9 / 12 Passed

Description

100%

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 that cleanly answers what the skill does and when to use it with concrete, natural trigger terms and a distinct niche. The trailing social-proof line ('Powers 300,000+ GPUs at xAI, AMD, NVIDIA, and LinkedIn') is mildly promotional but does not undermine the otherwise precise guidance.

Dimension	Reasoning	Score
Specificity	Lists multiple concrete actions — 'structured generation and serving', 'JSON/regex outputs, constrained decoding, agentic workflows with tool calls' — matching the score-3 anchor of several specific concrete actions rather than a single domain+action.	3 / 3
Completeness	Explicitly states both what ('Fast structured generation and serving for LLMs with RadixAttention prefix caching') and when via an explicit 'Use for ...' trigger clause, satisfying the score-3 anchor.	3 / 3
Trigger Term Quality	Covers natural developer terms users would say — 'JSON/regex outputs', 'tool calls', 'agentic workflows', 'faster inference than vLLM' — with good breadth of common variations.	3 / 3
Distinctiveness Conflict Risk	A clear niche (SGLang + RadixAttention prefix caching) with distinct triggers (constrained decoding, agentic tool calls, faster-than-vLLM prefix sharing) unlikely to fire for unrelated skills.	3 / 3
	Total	12 / 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/sglang/SKILL.md
Commit: 773a529

Reviewed: about 6 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.