skill-benchmarking

skill-benchmarking is a skill for Claude Code from rusel95/ios-agent-skills. It costs 149 tokens per session (4,736 once invoked), scanned A, original, MIT.

A benchmark runner for testing AI-agent skills against evaluation cases stored in an evals.json file. It works with coding assistants such as Claude Code, Gemini CLI, GitHub Copilot, Cursor, and Windsurf.

In plain words
What is it for?
Use it to run skill evaluations, compare models, produce benchmark JSON files, and re-grade previous benchmark runs.
Why use it?
It separates results caused by a skill from ordinary model behavior by comparing responses with and without that skill. It removes test assertions that do not distinguish between the two.

Skill for Claude Code

Written for Claude Code: shipped in a Claude Code plugin. Also seen: mentions subagents; mentions Claude Code; mentions Gemini CLI.

Part of the ios-agent-skills plugin — 12 skills shipped together

Good fit Use it to run skill evaluations, compare models, produce benchmark JSON files, and re-grade previous benchmark runs.

Compare 6 skills from other repositories ↓
Install with agentmods
npx agentmods add skills/rusel95/ios-agent-skills/benchmarking
Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

Any agent
npx skills add rusel95/ios-agent-skills --skill benchmarking
Clone the repo
git clone --depth 1 https://github.com/rusel95/ios-agent-skills

Made for: Claude Code.

Or install ios-agent-skills, the plugin that ships this one along with the rest of its 12 skills.

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 skill-benchmarking

README.md
[![agentmods](https://agentmods.dev/badge/skills/rusel95/ios-agent-skills/benchmarking/github.svg)](https://agentmods.dev/skills/rusel95/ios-agent-skills/benchmarking)
Your own site
<a href="https://agentmods.dev/skills/rusel95/ios-agent-skills/benchmarking"><img src="https://agentmods.dev/badge/skills/rusel95/ios-agent-skills/benchmarking/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 skill-benchmarking

Your own site · 80×15
<a href="https://agentmods.dev/skills/rusel95/ios-agent-skills/benchmarking"><img src="https://agentmods.dev/badge/skills/rusel95/ios-agent-skills/benchmarking.svg" alt="Reviewed on agentmods" width="80" height="20"></a>
Per session 149 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 4,736 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.
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 $0.00149 $0.04736
Opus 5 $0.00075 $0.02368
Sonnet 5 $0.00030 $0.00947
Haiku 4.5 $0.00015 $0.00474

Measured 8d ago against content hash 5c81e0fcf2e9, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-08, from the pricing page.

Security

Grade A, and why

skill-benchmarking 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 8d ago.

The scan reads SKILL.md. This mod also ships 9 executable files (__init__.py, aggregate_benchmark.py, aggregate.py, …), listed below but not scanned — reading those needs a real analyzer, not pattern matching.

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.

scripts/benchmarking/SKILL.md · 453 lines

How it starts

The opening of the file, as written. The whole thing — 453 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Skill Benchmarking

Strict, agent-agnostic benchmark runner for evals.json skill evaluation. Produces benchmark-<model>.json with pass rates and a discriminating assertion list. Only assertions that actually discriminate between with-skill and without-skill responses are kept; non-discriminating noise is removed via the assertion hygiene process.

This skill works with any AI coding assistant -- Claude Code, Gemini CLI, GitHub Copilot, Cursor, Windsurf, or any agent that can read files and run shell commands.


Quick Start for Non-Claude Agents

If you are using Gemini CLI, GitHub Copilot, Cursor, or another AI coding assistant:

  1. Read this file (scripts/benchmarking/SKILL.md) -- it is the complete workflow guide
  2. Follow the phases below in order. Each phase tells you exactly what to do
  3. Run Python scripts via your terminal or shell tool. All scripts use only the Python standard library (no pip installs needed)
  4. For grading (Phase 3), you MUST use a separate/fresh context that has NOT read the skill being tested. If your agent supports subagents or separate chat sessions, use that. If not, start a new chat session for grading
  5. File paths in this guide are relative to the repository root. Adjust if your working directory differs

Key differences from Claude Code usage

Claude Code feature Equivalent for other agents
Explore subagent Start a fresh chat session, or use your agent's subprocess/tool-use feature
Read / Write tools Use your agent's file read/write commands, or cat / shell redirects
Bash tool Use your agent's terminal/shell execution capability

All Python scripts are standalone and require only Python 3.10+. No external dependencies.


Non-Negotiable Invariants

  1. Grader isolation -- the context/session that generated responses does NOT grade them
  2. Evidence-only -- assertions pass only when the required content is EXPLICITLY stated in the response; implication, adjacency, and partial coverage all fail
  3. Blind grading -- the grader does not know whether it is grading a with_skill or without_skill response
  4. Model-agnostic -- model slug is always supplied by the caller; never infer or hardcode it

Read the full file on GitHub · 453 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. 8d ago First seen · 453 lines · 149 tokens per session scan A 5c81e0fcf2e9

Subscribe to this mod's changes

skill-benchmarking is a skill published in the GitHub repository rusel95/ios-agent-skills (7 stars, last pushed 4mo ago), licensed MIT. It adds 149 tokens to every session and 4,736 once invoked, about $0.0007 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 2026-08-31.

Related

Other skills, from other repositories

pdlc-feature

An automated workflow for taking a new software feature from a description or requirements document through requirements, design, testing, implementation, review, and release preparation. It keeps one feature ID across the documents and requires tests before implementation.

kanfu-panda/pdlc-skills · 30 tokens

pdlc-quality

A quality gate that runs configured coverage, end-to-end, and lint checks, compares their real results with declared targets, and creates a report for human approval.

kanfu-panda/pdlc-skills · 0 tokens

pdlc-test-setup

A project setup workflow that finds the programming language, package manager, test framework, and available checking tools, then records verified commands in test-commands.yml. It can also create test folders and connect local hooks.

kanfu-panda/pdlc-skills · 0 tokens

pdlc-tdd

A test-first development workflow guided by existing design documents. TDD means writing tests before the code they check; PDLC is a staged product-development process with required preparation steps.

kanfu-panda/pdlc-skills · 25 tokens

pdlc-implement

A guarded coding workflow that implements a feature from its design document and existing tests, then checks the result. It requires tests to be written first, following TDD (test-driven development).

kanfu-panda/pdlc-skills · 29 tokens

pdlc-api-mock

A command that reads API design documents and creates realistic sample responses and configuration for a mock service. A mock service imitates an API so frontend code can be tested before the real backend is ready.

kanfu-panda/pdlc-skills · 11 tokens