llm-evaluation-system: Skill for Claude Code

.claude/skills/port-a-benchmark/SKILL.md

port-a-benchmark is a skill for Claude Code from awslabs/llm-evaluation-system. It costs 118 tokens per session (1,902 once invoked), scanned A, original, Apache-2.0.

A procedure for importing an existing benchmark or evaluation into Inspect AI while preserving its original measurement. A benchmark is a repeatable test used to compare systems, and Inspect AI is a framework for running such evaluations.

In plain words
What is it for?
Use it when adding an outside benchmark, evaluation harness, paper implementation, or colleague's test suite to the repository.
Why use it?
It prevents small changes to datasets, prompts, or scoring rules from producing results that no longer mean the same thing as the original.

Skill for Claude Code ✓ vendor

Written for Claude Code: installed under .claude/. Also seen: mentions CLAUDE.md.

This is awslabs/llm-evaluation-system's own configuration. It tells Claude Code how to work on llm-evaluation-system itself, so it is not a mod to install elsewhere. Copy it as a starting point and replace the rules that are about this project. Everything llm-evaluation-system configures →

About the project

LLM Evaluation System is an agent-guided platform for evaluating language models and agents, generating datasets and configuring multiple judges from natural-language requests before producing an analysis report. It is for comparing model responses, testing agents, and creating document-grounded evaluation data.

awslabs/llm-evaluation-system · 23 stars · on GitHub

Reuse

Borrowing it

Nothing to install: this file belongs to awslabs/llm-evaluation-system. Take a copy, put it at the same path in your own repository, and replace the rules that are about this project with yours.

Copy the file
curl -O https://raw.githubusercontent.com/awslabs/llm-evaluation-system/main/.claude/skills/port-a-benchmark/SKILL.md
Clone the repo
git clone --depth 1 https://github.com/awslabs/llm-evaluation-system

Made for: Claude Code.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for port-a-benchmark

README.md
[![agentmods](https://agentmods.dev/badge/skills/awslabs/llm-evaluation-system/port-a-benchmark/github.svg)](https://agentmods.dev/skills/awslabs/llm-evaluation-system/port-a-benchmark)
Your own site
<a href="https://agentmods.dev/skills/awslabs/llm-evaluation-system/port-a-benchmark"><img src="https://agentmods.dev/badge/skills/awslabs/llm-evaluation-system/port-a-benchmark/github.svg" alt="Measured on agentmods" height="20"></a>

Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.

agentmods 80×15 button for port-a-benchmark

Your own site · 80×15
<a href="https://agentmods.dev/skills/awslabs/llm-evaluation-system/port-a-benchmark"><img src="https://agentmods.dev/badge/skills/awslabs/llm-evaluation-system/port-a-benchmark.svg" alt="Reviewed on agentmods" width="80" height="20"></a>
Per session 118 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 1,902 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. A grade says what 26 rules found in the file — not that it is safe. Third-party audits
  • NVIDIA SkillSpector pass 7 Sept 2026
How audits are shown
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5.1 $0.00118 $0.01902
Opus 5 $0.00059 $0.00951
Sonnet 5 $0.00024 $0.00380
Haiku 4.5 $0.00012 $0.00190

Measured 10d ago against content hash 0d50bda43d74, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-10, from the pricing page.

Security

Grade A, and why

port-a-benchmark scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 10d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

.claude/skills/port-a-benchmark/SKILL.md · 168 lines

How it starts

The opening of the file, as written. The whole thing — 168 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Port a Benchmark

A ported benchmark has exactly one job: produce the same measurement the original produces. A port that runs cleanly but measures something subtly different is worse than no port, because the numbers look authoritative and nobody can tell they're wrong.

The failure mode is specific and quiet: the dataset gets copied faithfully (obviously data), while a judge rubric living in a Python string literal gets "adapted" (looks like code). The suite stays green, because the suite exercises the plumbing. Only a line-by-line comparison against upstream catches it.

The rule

Copy the measuring instrument verbatim. Adapt only the plumbing.

Sort every file in the upstream repo into one of two buckets before writing any code. When unsure which bucket something belongs in, it goes in COPY.

COPY — byte-for-byte, no edits, no reformatting

Anything that influences what a score means:

  • Datasets, questions, golden/reference answers, expected outputs
  • Prompts of every kind — system prompts, judge/grader rubrics, user-turn scaffolding, few-shot examples, instruction preambles
  • Tool/function schemas (names, descriptions, required-ness all change model behaviour)
  • Scoring thresholds, weights, metric formulas, aggregation rules
  • Stop conditions, retry/recovery policy, turn boundaries

Prompts are data, not code. This is the trap. A prompt in a .py string literal looks like code, so it feels adaptable. It isn't. The wording is the instrument. Rewriting it for concision or clarity is not a refactor — it's recalibrating a scale nobody asked you to recalibrate.

ADAPT — rewrite freely

The mechanics of execution, which can't change what's measured:

  • Their runner → an Inspect @task with a solver + scorer
  • Their HTTP/SDK client → get_model() / execute_tools()
  • Their storage/logging → our log dir, viewer, S3 sync
  • Their CLI → an MCP tool
  • Their concurrency, retries-on-transport-error, progress output

Workflow

1. Read the whole upstream harness before writing anything

Read the full file on GitHub · 168 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 10d ago First seen · 168 lines · 118 tokens per session scan A 0d50bda43d74

Subscribe to this mod's changes

port-a-benchmark is a skill published in the GitHub repository awslabs/llm-evaluation-system (23 stars, last pushed today), licensed Apache-2.0. It adds 118 tokens to every session and 1,902 once invoked, about $0.0006 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.