skills.rest
Peterson-Benhame
Peterson-BenhameCOMMUNITY
Explore all skills by this author →
GitHub

harness-eval

Audits agent harness files for broken paths, redundant instructions, and low-usefulness surfaces.

Updated Sep 15, 2026
Search Tags:#audit#agents.md#redundancy#agent harness#instruction quality#llm judge#usefulness
One-click install
npx skills add https://github.com/Peterson-Benhame/agent-skills --skill harness-eval-peterson-benhame
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: harness-eval
Source: https://github.com/Peterson-Benhame/agent-skills/tree/main/packages/skills-catalog/skills/%28development%29/harness-eval
Command: npx skills add https://github.com/Peterson-Benhame/agent-skills --skill harness-eval-peterson-benhame

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) and references (resource) components.

What problem does it solve? Agent harnesses (AGENTS.md, rules files, skills, and referenced docs) accumulate broken path citations, redundant instructions, and theory-heavy content that wastes context tokens without changing agent behavior. This Skill evaluates a repository's harness across three tracks and produces report-only findings without auto-editing anything. ## Core Features & Use Cases - Track A — Correctness: Deterministic script checks that cited paths and commands actually exist, preferring false negatives over false BROKEN flags. - Track B — Redundancy: Dual blind LLM judges with planted calibration traps classify claims as Ship/Review/Hold based on rediscovery cost. - Track C — Usefulness: Dual judges plus a fan-in gate classify surfaces as Slim/Keep-core/Mixed/Hold, emitting a mechanical KEEP/CUT apply plan for Mixed surfaces. - Use Case: Run a full harness eval on a repo to find dead links in AGENTS.md, trim instructions an agent would rediscover anyway, and identify skill files that are mostly theory before approving any edits. ## Quick Start Ask the agent to run a harness eval on this repository, then answer the Q1 optional-docs and Q2 track-budget questionnaires to choose the evaluation scope.

Frequently Asked Questions about harness-eval

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I audit AGENTS.md and agent skills for broken paths?▼

Run the harness eval inventory script followed by Track A, which deterministically checks that every cited path and command exists in the repo. It prefers false negatives over false BROKEN flags and skips placeholders like SPEC_FOLDER or template variables.

What is the difference between Track B redundancy and Track C usefulness?▼

Track B asks whether an agent would cheaply rediscover an instruction without the harness text, producing Ship/Review/Hold bands. Track C asks whether deleting a surface would change agent behavior, producing Slim/Keep-core/Mixed/Hold bands. Ship does not equal Slim.

Does harness-eval automatically edit my AGENTS.md or skills?▼

No, it is report-only by default. Judgment is separated from remediation, and edits to Ship, Slim, or Mixed surfaces happen only when the user explicitly asks after reviewing the agreement reports.

Are ADRs and RFCs included in the harness evaluation?▼

No, ADR and RFC decision-record trees are always excluded from T2 usefulness surfaces. Other cited project docs are opt-in only after the user approves them via the Q1 questionnaire.

Why did the trap gate fail during a redundancy or usefulness run?▼

Trap gate FAIL means the blind judge misclassified planted calibration claims, so Ship or Slim bands cannot be trusted. Fix the plants so they are not verbatim deck duplicates, rescore the plant rows, and re-run the merge script.

When should I not use harness-eval?▼

Do not use it for harness setup or initialization, feature spec-driven work, or applying Ship/Slim trims unless the user explicitly requests edits after reviewing the report. It is an audit tool, not a scaffolding or remediation tool.

Related Skills

Explore More Software Engineering Skills→
masermediagroup-stackmasermediagroup-stack
1

migrate-to-skills

Convert Cursor rules and slash commands into SKILL.md directories.

Community
Intermediate
tech-leads-clubtech-leads-club
5.1k

harness-eval

Audit agent harness files for broken paths, redundant instructions, and low-usefulness surfaces.

Official
Advanced
jorgealvesjorgealves
2

prompt-injection-scanner

Audit agent skill instructions and system prompts for prompt injection vulnerabilities.

Community
Advanced
SavigeSystemZSavigeSystemZ

etherweave-code-review

Automate code review against Etherweave PROJECT_RULES and architecture boundaries.

Official
Advanced