How to use the Reference Originality Audit agent skill

What is an agent skill?

An agent skill is a folder of instructions and supporting files that an AI agent loads when a task matches the skill's description. This skill's folder contains SKILL.md, a rubric file, a Python helper script and an agents/openai.yaml file.

The skill does not run by itself. The agent uses it when your request matches what it is for, such as comparing a site with its references.

What the skill makes the agent do

The agent builds a registry of the materials you supply, inventories the current site and its history, and compares it category by category: text, brands, numbers, images, assets, videos, structure and motion, and history. Every candidate red flag must cite an exact location on the current site and an exact location in the reference material.

Findings are called originality risks, overlaps or red flags. The skill is written to avoid declaring legal plagiarism from visual similarity alone, and it separates facts from inferences. Where evidence is missing, the report says so instead of marking the area as passed.

The report opens with one verdict, such as Clear in checked scope, Changes recommended or Block release, followed by a red-flag table ordered by severity, a category pass list, history findings, access gaps and a prioritized fix plan.

• Compares current output and repository history against your references
• Rates findings by severity and explains the evidence for each
• Proposes concrete fixes such as rewriting copy or replacing images
• Lists what it could not verify

Install the skill

Download the ZIP for this skill from our page and unzip it. For a personal install, place the folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder.

In the Claude apps, you can instead upload the skill under Settings → Capabilities → Skills.

• Personal: unzip into ~/.claude/skills/
• Project: unzip into the project's .claude/skills/
• Claude apps: Settings → Capabilities → Skills, then upload

Prepare your materials

Before you ask for an audit, gather local copies of the site and the reference materials. The skill is designed to work against a supplied reference set, so an audit with no references is not what it is for.

Local evidence captured when the brief was written is preferred. Live reference URLs are optional and are used only to fill gaps, because pages may have changed.

• Local copies of the site source and, if possible, built output
• Reference pages, capture packs, screenshots, copy, images, videos or code
• Git history for the repository, if you want history searches

Trigger the skill with natural requests

You do not need special syntax. Describe the job and point the agent at your files. Mention the references explicitly so the agent knows what to compare against.

The skill's description covers requests to compare site output with reference pages, capture packs and copy, and to raise red flags and propose fixes.

• Check this new client site against the references in ./refs for copied copy or images before launch.
• Audit the current homepage and our git history for assets copied from the inspiration pages.
• Compare the built site with the reference screenshots and list red flags by severity, with fixes.

Run the evidence helper (optional)

The bundled script scripts/build_evidence_inventory.py scans local files and finds exact-file matches, suspicious basename reuse, normalized text overlap and repeated number tokens, including historical matches. It needs Python 3. Its output is a set of leads for human review, not a verdict. The agent fills in the command with real paths; the skill's template uses placeholders for the site root, reference paths and an output file.

• Requires Python 3
• Writes results to a JSON file you choose
• Uses local files only; the skill lists no network use

Read the result and act on it

Start with the verdict, then the red-flag table. Each row should show the current-site location, the reference location, what overlaps and what differs. Blockers and high-severity items are the ones to fix first.

Fixes are meant to replace the copied element. Recoloring alone is not treated as a fix for copied identity, copy, media or composition. Check the report's list of unverified areas before you treat a result as complete.

• Verdict: Clear in checked scope, Clear with low-risk similarities, Changes recommended, Block release or Blocked by missing evidence
• Fix plan: prioritized, with replacements for copied text, images, assets, videos or structure