CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

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

78%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Low

Low-risk findings worth noting

Fix and improve this skill with Tessl

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

Quality

Content

61%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.

The body is a well-structured, highly actionable usage guide with real, executable examples and properly linked one-level-deep references. Its weaknesses are redundancy (benchmark/speedup claims presented three ways), absent validation checkpoints in the launch-then-query workflow, and reference-scope material inlined in the body.

Suggestions

Consolidate the repeated speedup claims into the single throughput table and move detailed benchmark numbers into references/radix-attention.md, dropping the Request 1/Request 2 illustration and the duplicated 'RadixAttention speedup' list.

Add validation checkpoints to the Quick start workflow: verify the server is healthy (e.g., curl the /health or /v1/models endpoint) before running client code, with brief error-recovery guidance for launch failures.

Move the supported-models, hardware-support, and benchmark sections into the existing reference files to keep SKILL.md as a lean overview.

DimensionReasoningScore

Conciseness

Mostly efficient code-and-command content, but the speedup claims are stated redundantly (inline in the RadixAttention example '→ 5× faster!', in the 'Performance benchmarks' list, and again in the throughput table), the OpenAI API section repeats the server-launch command already shown, and the Request 1/Request 2 illustration is padding. Not a 4 because multiple sections could be trimmed without losing information.

3 / 5

Actionability

Nearly all guidance is concrete, runnable code: install commands, `python -m sglang.launch_server` invocations, `@sgl.function` examples covering JSON/regex/grammar constraints, batching, vision, and the OpenAI-compatible API. Minor gaps keep it below 5: the grammar example references undefined terminals (NAME, NEWLINE, INDENT) and is never run, and `simple_gen.bind` depends on an earlier example.

4 / 5

Workflow Clarity

The Quick start implies a clear sequence (install → launch server → run inference) but there are no validation checkpoints: nothing verifies the server is up before running client code, and there is no error-recovery guidance for launch or connection failures. This matches anchor 3 (sequence present, checkpoints missing/implicit); not a 4 because even a basic health-check step is absent.

3 / 5

Progressive Disclosure

Good structure with a clearly signaled References section pointing to three real, one-level-deep files (references/structured-generation.md, references/radix-attention.md, references/deployment.md) with descriptive labels. Below 5 because the in-body 'RadixAttention (Key Innovation)' and 'Performance benchmarks' sections duplicate the radix-attention reference's stated scope ('How it works, optimization, benchmarks'), and the supported-models/hardware lists are reference material inlined in the body.

4 / 5

Total

14

/

20

Passed

Description

88%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 explicitly states both what the skill does and when to use it, with concrete capability enumeration and natural trigger phrases. Its only weaknesses are a couple of missing natural synonyms and slightly generic performance-related triggers that could overlap with other inference-serving skills.

DimensionReasoningScore

Specificity

Concrete actions are enumerated: 'structured generation and serving for LLMs', 'RadixAttention prefix caching', 'JSON/regex outputs, constrained decoding, agentic workflows with tool calls'. Multiple specific capabilities with broad coverage of the tool's main functions; not a 4 because the gaps (e.g., speculative decoding, vision models) are peripheral to what the skill is triggered for.

5 / 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. Matches the anchor-5 example's structure.

5 / 5

Trigger Term Quality

Natural trigger phrases are present: 'JSON/regex outputs', 'agentic workflows with tool calls', 'faster inference than vLLM', 'prefix sharing'. A few natural variations users might say are missing ('structured output', 'LLM serving', 'grammar-constrained output'), so it sits between anchors 4 and 5 but noticeably above the midpoint.

4 / 5

Distinctiveness Conflict Risk

Named technology (RadixAttention, SGLang, explicit vLLM comparison) gives it a clear niche, but generic triggers like 'faster inference' and 'serving for LLMs' create minor overlap risk with other inference-serving skills (vLLM, TensorRT-LLM). Not a 5 because those broad phrases could pull in closely related serving skills.

4 / 5

Total

18

/

20

Passed

Validation

87%

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

Validation — 14 / 16 Passed

Validation for skill structure

CriteriaDescriptionResult

metadata_version

'metadata.version' is missing

Warning

frontmatter_unknown_keys

Unknown frontmatter key(s) found; consider removing or moving to metadata

Warning

Total

14

/

16

Passed

Repository

synthetic-sciences/openscience

Path

backend/cli/skills/ml-inference/sglang/SKILL.md

Reviewed

Table of Contents

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.