instruction-eval

A tool for testing whether one change to an agent’s operating conditions affects its answers. It runs the same prompts with and without a changed instruction, reference, tool, permission, hook, or skill file, then creates an HTML comparison report.

In plain words
What is it for?
Use it for instruction experiments, such as comparing two rule wordings, checking whether reference material is being read, or measuring the effect of adding or removing an agent condition.
Why use it?
It provides repeated before-and-after evidence when deciding whether a rule changes behavior or can be removed. It is not for testing application code.

Plugin for Claude Code

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.

Claude Code
/plugin marketplace add 2ykwang/agent-skills
agentmods
npx agentmods add plugins/2ykwang/agent-skills/instruction-eval
Clone the repo
git clone --depth 1 https://github.com/2ykwang/agent-skills

Made for: Claude Code.

Per session not measured What this adds to a session before it is invoked.
When invoked not measured Not applicable: nothing here is loaded into a session.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Security

Grade A, and why

instruction-eval 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 2d 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.

plugins/instruction-eval/.claude-plugin/plugin.json · 13 lines

What it actually says

{
  "name": "instruction-eval",
  "version": "0.5.1",
  "description": "Change one condition an agent runs under (instruction text, the reference material instructions point at, MCP tools and permissions, hooks, skill files), then run the same prompts before and after, several times each, to see what actually changes. Produces an HTML report showing both arms' answers side by side. Use when deciding whether to add a rule, when you need evidence that an existing rule can be deleted, when picking between two wordings of the same rule, or when checking whether material you placed is actually being read. Triggers on 'does this rule do anything', 'compare before and after adding this', 'can I drop this', 'difference with and without this', 'run an A/B'. Not for application code changes or anything automated tests already verify. Built on Claude Code: it spawns `claude -p` subprocesses to run both arms.",
  "author": {
    "name": "2ykwang",
    "url": "https://github.com/2ykwang"
  },
  "repository": "https://github.com/2ykwang/agent-skills",
  "homepage": "https://github.com/2ykwang/agent-skills/tree/main/skills/instruction-eval",
  "license": "MIT"
}
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. 2d ago First seen · 13 lines scan A 72e9acd7a547

Subscribe to this mod's changes

instruction-eval is a plugin published in the GitHub repository 2ykwang/agent-skills (18 stars, last pushed 29d ago), licensed MIT. Its token cost is not measured: this kind of file is read by the harness, not the model. 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.