CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

jetson-llm-benchmark

Benchmark Jetson LLM/VLM serving performance across vLLM, llama.cpp, and Ollama with structured JSON output.

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 ./skills/jetson-llm-benchmark/SKILL.md
SKILL.md
Quality
Evals
Security

Quality

Content

81%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, well-sequenced script-based skill with strong error-recovery feedback loops and verified bundle references. Its main weaknesses are instruction duplication across the Instructions/Expected Workflow/Limitations sections and keeping all per-wrapper detail inline in SKILL.md rather than splitting some of it into reference files.

Suggestions

Consolidate the repeated '--help first' directive and the vLLM container-selection detail (currently duplicated nearly verbatim in section A and 'Limitations') into a single location and cross-reference it.

Move the three per-wrapper JSON output-contract schemas into a references/output-contract.md file (or per-wrapper reference files) and keep only the shared envelope in SKILL.md to reduce inline bulk.

Merge the overlapping guidance in 'Instructions', 'Expected Workflow', and 'Three paths' into one decision-tree section so the pick-one-wrapper rule is stated once.

DimensionReasoningScore

Conciseness

The body is mostly high-value non-obvious Jetson knowledge (container selection, comparability caveats, exit codes), but it repeats instructions across sections: the '--help first' directive appears in both 'Instructions' and 'Expected Workflow', and the vLLM benchmark container selection ('vllm/vllm-openai:latest on Thor and Orin JetPack 7.2 / L4T r39+... ghcr.io/nvidia-ai-iot/vllm:latest-jetson-orin on older Orin') is stated nearly verbatim in both section A and 'Limitations'. This matches the 3 anchor ('mostly efficient but some unnecessary explanation or could be tightened') — not a 2, since no space is spent on concepts Claude already knows, and not a 4, since the duplication is substantive rather than minor.

3 / 5

Actionability

Every runtime path ships a copy-paste-ready command with concrete flags ('scripts/bench_vllm.sh --model <hf-repo-id-being-served> --concurrency 1,8 --input-len 2048 --output-len 128 --num-prompts 50'), a --help verification step, a complete JSON output contract per wrapper, and exit-code-driven recovery guidance (exit 2/3, Docker stderr). This fully matches the 5 anchor ('fully executable; copy-paste ready commands; specific examples cover the common cases').

5 / 5

Workflow Clarity

The workflow is clearly sequenced — prerequisites, pick exactly one wrapper by named runtime, invoke with --help before composing the answer, run the benchmark, interpret the JSON envelope and warnings, then hand off — and the 'Error Handling' section provides explicit feedback loops (exit 2 → re-run with --help and correct the path; exit 3 → start the service, pull the model, or pass --container; malformed JSON → preserve the raw error and rerun). This matches the 5 anchor ('clear sequence with explicit validation steps; feedback loops for error recovery'), and the operation is measurement rather than destructive/batch, so no cap applies.

5 / 5

Progressive Disclosure

The bundle structure is sound: the 'Available Scripts' table plus inline references point to all three real script files (scripts/bench_vllm.sh, scripts/bench_llama_cpp.sh, scripts/bench_ollama.sh — all verified present), references are one level deep, and sections are clearly headed. It is not a 5 because all detail lives in SKILL.md itself — the three per-wrapper JSON output schemas and the duplicated container-selection notes are inline content that could be split into reference files, and the {baseDir} invocation guidance is spread across multiple sections. It is above a 3 because navigation is easy and nothing is buried or nested.

4 / 5

Total

17

/

20

Passed

Description

70%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 concrete, distinctive description that clearly states what the skill does and on which runtimes, but it omits any explicit 'use when' trigger guidance, which both caps completeness and leaves natural trigger terms like latency/throughput/measure unrepresented.

Suggestions

Add an explicit trigger clause, e.g. 'Use when the user asks to benchmark, measure, or compare LLM/VLM latency or throughput on a Jetson device, or to quantify a serving deployment before/after tuning.'

Include natural user phrasings and synonyms such as 'latency', 'throughput', 'tokens per second', 'TTFT', or 'measure serving performance' so the description matches how users actually ask.

Mention the comparison use case (before/after flag or power-mode changes) in the description to make the 'when' component concrete.

DimensionReasoningScore

Specificity

The description names a concrete action ('Benchmark Jetson LLM/VLM serving performance'), enumerates three specific runtimes ('across vLLM, llama.cpp, and Ollama'), and specifies the output format ('with structured JSON output'). It falls short of the 5 anchor because it covers a single benchmark action rather than multiple distinct concrete actions, and it is well above the 3 anchor since the scope is precise rather than merely naming a domain.

4 / 5

Completeness

The 'what' is clearly answered (benchmark serving performance across three runtimes with JSON output), but there is no 'Use when...' clause or equivalent explicit trigger guidance, which per the judging guidelines caps completeness at 3 ('has a clear what but when is missing or only weakly implied'). It is not a 2 because the 'what' is concrete and specific, and not a 4 because no 'when' guidance of any kind is present.

3 / 5

Trigger Term Quality

Natural keywords include 'Benchmark', 'LLM', 'VLM', 'vLLM', 'llama.cpp', 'Ollama', and 'Jetson', which a user on this hardware would plausibly say. Common variations like 'latency', 'throughput', 'measure', or 'compare models' are missing, matching the 4 anchor ('good keyword coverage; a few natural terms missing') rather than the comprehensive synonym coverage of the 5 anchor.

4 / 5

Distinctiveness Conflict Risk

The Jetson hardware qualifier plus the three named runtimes carve out a clear niche with minimal conflict risk against generic benchmarking or serving skills; the trigger terms (Jetson, vLLM, llama.cpp, Ollama) are distinctive. It is not a 4 because the overlap with any closely related skill is negligible — a generic LLM-benchmark skill would not claim the Jetson-specific serving-performance scope.

5 / 5

Total

16

/

20

Passed

Validation

81%

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

CriteriaDescriptionResult

metadata_version

'metadata.version' is missing

Warning

metadata_field

'metadata' should map string keys to string values

Warning

frontmatter_unknown_keys

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

Warning

Total

13

/

16

Passed

Repository

NVIDIA/skills

Path

skills/jetson-llm-benchmark/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.