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

79%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Passed

No findings from the security scan

Fix and improve this skill with Tessl

tessl review fix ./backend/cli/skills/ml-inference/sglang/SKILL.md

Quality

Content

75%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.

A highly actionable reference packed with executable, copy-paste-ready examples and clean quick-start sequencing, supported by real one-level-deep reference files. Its weakness is conciseness: performance/scale claims recur across multiple sections and some inlined material (benchmarks, model/hardware lists) duplicates what belongs in the reference files.

Suggestions

Collapse the repeated performance claims into one place: keep the throughput table and drop the prose 'Performance benchmarks' bullet list and the duplicated '5×/10×' lines from the RadixAttention section, or move all benchmarks into references/radix-attention.md.

Trim or move the 'Supported models' and 'Hardware support' lists into references/deployment.md, keeping only a one-line summary plus a link in SKILL.md.

Add a one-step validation checkpoint after 'Launch server' (e.g. curl the /v1/models endpoint or check logs for 'ready') so the quick-start sequence confirms the server is up before running inference.

Dimension	Reasoning	Score
Conciseness	The body is mostly efficient concrete code, but performance claims are repeated across the RadixAttention section ('5× faster', '10× faster', 'Zero configuration'), a prose 'Performance benchmarks' section, and a throughput table, and the benchmark prose reads as padded marketing rather than reference material.	3 / 5
Actionability	Copy-paste-ready executable code throughout — install commands, server launch flags, sgl.function programs for JSON/regex/grammar, agent tool-calling, multi-turn chat, OpenAI-compatible curl — covering the common serving and structured-generation cases.	5 / 5
Workflow Clarity	Quick start sequences install → launch server → basic inference → structured output clearly, and the skill targets non-destructive serving so the missing-validation cap does not apply; minor gaps only, with no explicit checkpoints for verifying a server came up healthy before use.	4 / 5
Progressive Disclosure	A dedicated References section links to three real one-level-deep bundle files (structured-generation.md, radix-attention.md, deployment.md) with clear markdown links, but the SKILL.md itself is long and inlines benchmark tables, supported-models, and hardware lists that overlap with the referenced deep-dives, leaving minor organization gaps versus a lean overview.	4 / 5
	Total	16 / 20 Passed

Description

83%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 strong description that cleanly states what SGLang does and when to reach for it, with explicit trigger phrasing and concrete capability terms. The main weakness is the trailing scale/social-proof sentence ('Powers 300,000+ GPUs at xAI, AMD, NVIDIA, and LinkedIn'), which is marketing padding rather than a capability or trigger and slightly dilutes specificity.

Dimension	Reasoning	Score
Specificity	Lists several concrete capabilities — 'structured generation and serving for LLMs', 'RadixAttention prefix caching', 'JSON/regex outputs', 'constrained decoding', 'agentic workflows with tool calls' — but the trailing 'Powers 300,000+ GPUs at xAI, AMD, NVIDIA, and LinkedIn' is marketing fluff rather than a capability, leaving minor gaps versus the comprehensive anchor.	4 / 5
Completeness	Explicitly answers both what ('Fast structured generation and serving for LLMs with RadixAttention prefix caching') and when ('Use for JSON/regex outputs, constrained decoding, agentic workflows with tool calls, or when you need 5× faster inference than vLLM with prefix sharing') with concrete trigger phrases.	5 / 5
Trigger Term Quality	Good natural keyword coverage ('JSON/regex outputs', 'agentic workflows with tool calls', 'faster inference than vLLM', 'prefix sharing'), but common synonyms like 'schema', 'function calling', and 'KV cache' are absent, so it stops short of comprehensive.	4 / 5
Distinctiveness Conflict Risk	SGLang has a clear niche (RadixAttention prefix caching + structured generation) with distinct triggers, but it explicitly names vLLM as a comparison, signaling real overlap risk with a closely related serving skill.	4 / 5
	Total	17 / 20 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: synthetic-sciences/openscience
Path: backend/cli/skills/ml-inference/sglang/SKILL.md
Commit: 3a6c3a9

Reviewed: 1 day 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.