skill-iterator

skill-iterator

A Claude Code plugin that puts a skill to work, scores it across three levels, and hands back a concrete list of improvements — without ever touching the skill it is testing.

What it does

Writing a skill is easy. Knowing whether it actually helps is not. skill-iterator answers that question empirically:

• Run it for real — make the skill perform an actual task, not a hypothetical one.
• Score in three layers — L1: did it produce anything? · L2: is the output correct? · L3: is it any good?
• Run every task twice — once with the skill, once without — so you can isolate what the skill itself contributed.
• Get an actionable scorecard — every suggestion names a file, a change, and a reason. No filler.

Four commands

The plugin splits by user intent — pick the one that matches what you want to know:

• /skill-iterator — umbrella router; start here if you are not sure which flow you need.
• /iterate-skill — L1/L2/L3 content scoring + improvement suggestions over multiple iterations.
• /trigger-tune — optimize a skill's description so it fires at the right time (trigger-rate tuning).
• /blind-test — end-to-end output test with anti-fixture-leak guards.

Design principles

• Read-only on the target. It never mutates the skill under test — it only emits a suggestions.json. Applying changes is a separate, deliberate step.
• With / without double-run. The core invariant: you cannot claim a skill helps unless you have measured the same task without it.
• Three-layer lint. Schema and structure are validated by a PostToolUse hook, a git pre-commit gate, and a manual validator.

Install

This repo is a Claude Code marketplace containing one plugin. From inside Claude Code:

# 1. add the marketplace
/plugin marketplace add synergy321/skill-iterator

# 2. install the plugin
/plugin install skill-iterator@skill-iterator-plugin

plugin	skill-iterator
marketplace	skill-iterator-plugin

Skills, agents, and the lint hook are auto-discovered — no extra configuration.

Requirements & honest caveats

• Read-only evaluation works out of the box. Scoring, blind-test, and the suggestion scorecard run with just Claude Code + Python 3.
• Applying suggestions expects an external editing workflow. Several skills route the "apply changes" step to skill-creator, an authoring tool that is not bundled with this plugin. Without it you still get every finding — you just apply the edits with your own workflow.
• Batch mode needs an API key. run_loop.py drives the Anthropic SDK directly and reads ANTHROPIC_API_KEY from the environment; the agent-spawn path does not.
• Skill prompts and output are in Chinese. The evaluation methodology is language-agnostic, but the user-facing text is Chinese.

License

MIT © 2026 Travis Chong. Use it, fork it, ship it.