carve-harness: Skill for Claude Code

.claude/skills/eval-goldenset/SKILL.md

eval-goldenset is a skill for Claude Code from claude-code-expert/carve-harness. It costs 105 tokens per session (4,685 once invoked), scanned A, original, MIT.

A fixed test set that scores an agent’s output against defined criteria and records how the scores change over time. It can also measure how reliably repeated attempts pass.

In plain words
What is it for?
Use it to check output quality after changes and track pass@k or pass^k, which describe success across repeated attempts.
Why use it?
It shows with numbers whether changes to prompts, agents, skills, or rules caused quality to get worse. A regression is a change that breaks something that previously worked.

Skill for Claude Code

Written for Claude Code: installed under .claude/. Also seen: reads .claude/ paths; mentions CLAUDE.md; mentions AGENTS.md.

This is claude-code-expert/carve-harness's own configuration. It tells Claude Code how to work on carve-harness 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 carve-harness configures →

Needs its repository: it runs a file that does not travel with it, so clone the repository first. The line is bash .claude/hooks/carve-validate.sh # specs/goldenset/*.json 구조 검증.

Reuse

Borrowing it

Nothing to install: this file belongs to claude-code-expert/carve-harness. 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/claude-code-expert/carve-harness/main/.claude/skills/eval-goldenset/SKILL.md
Clone the repo
git clone --depth 1 https://github.com/claude-code-expert/carve-harness

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 eval-goldenset

README.md
[![agentmods](https://agentmods.dev/badge/skills/claude-code-expert/carve-harness/eval-goldenset/github.svg)](https://agentmods.dev/skills/claude-code-expert/carve-harness/eval-goldenset)
Your own site
<a href="https://agentmods.dev/skills/claude-code-expert/carve-harness/eval-goldenset"><img src="https://agentmods.dev/badge/skills/claude-code-expert/carve-harness/eval-goldenset/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 eval-goldenset

Your own site · 80×15
<a href="https://agentmods.dev/skills/claude-code-expert/carve-harness/eval-goldenset"><img src="https://agentmods.dev/badge/skills/claude-code-expert/carve-harness/eval-goldenset.svg" alt="Reviewed on agentmods" width="80" height="20"></a>
Per session 105 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 4,685 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 warn 7 Sept 2026
SkillSpector: 1 finding, up to medium

These are SkillSpector’s own severities. On a checked sample its high-severity flags on skills were ~96% false positives — a documented command, a public API, a “never do X” rule — so we show them as a caution to read, not a verdict. Why →

  • medium Agent Snooping · line 203
    Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.
    Fix: Remove all code or instructions that list or read other skills' files or directories. Skills should operate independently; cross-skill access is a privilege escalation.
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 $0.00105 $0.04685
Opus 5 $0.00053 $0.02343
Sonnet 5 $0.00021 $0.00937
Haiku 4.5 $0.00011 $0.00468

Measured 3d ago against content hash 9a397202185a, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-09, from the pricing page.

Security

Grade A, and why

eval-goldenset 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 3d 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/eval-goldenset/SKILL.md · 204 lines

How it starts

The opening of the file, as written. The whole thing — 204 lines — stays where its author put it; the contents beside it link to each section on GitHub.

eval-goldenset — 골든셋 정량평가 (모델 무관 SOP)

권위는 도구가 아니라 이 절차에 있다. carve-eval 워크플로는 이 절차를 코드로 자동 실행할 뿐이다. 태스크당 완성도는 checklist-loop(항목 5축 채점)이 담당하고, 이 스킬은 고정 케이스 집합의 품질을 시간축으로 관리한다.

첫 골든셋을 만드는 중이면 /eval-init을 써라 — 이 문서의 절차를 프로젝트 분석 + 대화형 인터뷰로 태우는 실행기다(SOP ⟷ 실행기 관계는 checklist-loopcarve-verify-loop와 같다). 이 문서는 형식·판정 기준의 정본이고, 셋업 이후의 증설(트레이스 마이닝)도 여기 절차를 따른다.

언제 쓰나

  • 프롬프트·CLAUDE.md·에이전트·스킬·규칙을 바꾼 뒤 회귀 확인 — "이 변경이 출력을 나쁘게 만들지 않았나"를 숫자로.
  • 데모는 되는데 프로덕션은 무너지는지 — pass@k(능력) vs pass^k(일관성) 분리 측정.
  • 안 쓰는 경우: 정답 없는 탐색·리서치, 골든셋을 유지할 사람이 없을 때(유지 안 하면 추이가 무의미).

골든셋 형식 (specs/goldenset/*.json)

{
  "suite": "cs-agent",
  "cases": [
    {
      "id": "refund-policy",
      "version": "1.0",
      "prompt": "고객: 어제 산 제품 환불돼요? 규정대로만 답하라.",
      "k": 3,
      "assert": [
        { "type": "contains",    "value": "14일" },
        { "type": "regex",       "value": "환불|반품" },
        { "type": "not_contains","value": "무조건" },
        { "type": "llm-rubric",  "value": "규정 수치를 정확히 인용하고 규정 외 약속을 하지 않는다" }
      ]
    }
  ]
}
  • assert.type 3계층:
    • 텍스트(순수 채점): contains · not_contains · regex · not_regex
    • 상태(.claude/hooks/eval-state.sh가 결정론 채점 — 워크디렉토리의 실제 상태): file_exists(경로) · file_contains(경로::리터럴) · cmd_exit0(명령 exit 0) · git_diff_contains(diff 내 리터럴) · log_contains(logs/*.jsonl::<jq 불리언 필터> — 워크디렉토리의 하네스 훅 로그로 경로(trajectory)를 결정론 채점. 예: .decision=="block" and .tool=="Write". 외부 target(claude·exec:)에서만 유효 — 세션 응답자는 로그가 리포 logs/로 간다)
    • 정성: llm-rubric(evaluator 위임 — "평가의 평가" 문제가 있으니 상태 assert로 대체 가능하면 대체)
  • setup(선택): 케이스 실행 전 격리 워크디렉토리에서 실행할 bash 스크립트(환경 구성 — 파일·git 초기화 등). 상태 assert 또는 setup이 있으면 respondent는 리포 밖 임시 디렉토리에서 실행된다(골든셋 정답 비노출).
  • k: 반복 실행 횟수(기본 1, 상한 10). k>1이면 pass@k·pass^k가 의미를 가진다.
  • tags(선택): ["required", "category:domain_safety"]. required 케이스는 하나라도 완전 green(100)이 아니면 eval-gate가 평균과 무관하게 regressed로 막는다(블루프린트 §6.5 관문 ①). category:*는 리포트 분류용. 골든셋에 required가 0건이면 carve-validate가 NOTE(--strict면 ERROR) — 안전·결제·인증 케이스에 붙여라. 이 리포는 harness-guard 5건이 required.
  • version(필수): 케이스 정의의 버전. 케이스를 고치면 반드시 올린다. 버전 없이 점수만 쌓으면 run #3과 run #7이 서로 다른 문제를 푼 점수인데도 같은 축에 그려져 추이가 조용히 무의미해진다. 추이 엔트리에 caseVersion으로 함께 기록되고, 직전 run과 다르면 [VERSION CHANGED]로 경고한다.
  • 알 수 없는 assert 타입·잘못된 정규식·상태 채점 불능은 전부 fail-closed(통과로 새지 않음).

Read the full file on GitHub · 204 lines

Files

What ships with it

3 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.

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. 3d ago Changed · +59 lines 9a397202185a
  2. 9d ago First seen · 145 lines · 105 tokens per session scan A 0c3ca85df791

Subscribe to this mod's changes

eval-goldenset is a skill published in the GitHub repository claude-code-expert/carve-harness (13 stars, last pushed yesterday), licensed MIT. It adds 105 tokens to every session and 4,685 once invoked, about $0.0005 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.

Related

Other skills, from other repositories

research-engineer

An uncompromising Academic Research Engineer. Operates with absolute scientific rigor, objective criticism, and zero flair. Focuses on theoretical correctness, formal verification, and optimal implementation across any required technology.

davila7/claude-code-templates · 43 tokens

tika-eval-compare

Compare extracts from two Tika builds over a corpus to detect regressions in content, encoding, exceptions, and embedded-document handling. Use for "compare before/after extracts", "eval this change against the corpus".

apache/tika · 50 tokens

neuron-evaluation-engineer

Create and run AI evaluations with datasets, assertions, and output drivers in Neuron AI. Use this skill whenever the user mentions evaluation, testing AI systems, creating evaluators, dataset-driven testing, assertion-based validation, or wants to measure AI system performance. Also trigger for tasks involving…

neuron-core/neuron-ai · 77 tokens

jetson-validate-image

Use after jetson-flash-image to run static BSP checks, on-target smoke/regression tests on a flashed DUT, or both. Not for build or flash steps. Triggers: validate bsp, on-target validation.

NVIDIA/skills · 50 tokens

atmos-validation

Validate Atmos projects, components, arbitrary JSON Schema inputs, EditorConfig, and GitHub Actions; use affected-file selection and native CI annotations.

cloudposse/atmos · 31 tokens

skill-benchmark

Benchmark AI skill effectiveness by measuring implementation quality against legacy constraints.

HoangNguyen0403/agent-skills-standard · 16 tokens