Global Greenwashing Claim Auditor

A portable agent skill for first-pass analysis of environmental marketing claims. It identifies potentially misleading or insufficiently substantiated claims and reports universal claim-construction risks separately from findings specific to Canada, the European Union, and the United Kingdom.

Use it to review product descriptions, advertising copy, sustainability reports, presentations, websites, social content, packaging text, and other public-facing communications before publication or legal review.

Supported inputs

• Pasted conversational text or a text file
• CSV datasets
• XLSX workbooks
• DOCX documents
• PPTX presentations
• Public HTTP or HTTPS website URLs

The bundled scripts use only Python’s standard library. XLSX, DOCX, and PPTX content is extracted directly from their Open XML files, so no additional Python packages are required.

Add the skill to an agent

1. Download and unpack the skill bundle.
2. Add the unpacked skill to the agent.
3. Ensure the uploaded skill retains its SKILL.md, scripts/, and references/ directories.
4. Ask the agent to audit content, specifying the intended markets and publication date when known.

The scripts use Python 3 and workspace-relative paths so they can run across supported agent environments.

Example requests

• “Audit this product description for greenwashing risks in Canada, the EU, and the UK.”
• “Review the attached sustainability claims spreadsheet and prioritize the highest-risk rows.”
• “Scan this presentation for environmental claims that need qualification or substantiation.”
• “Check this public product page for vague, absolute, carbon-neutral, and recyclability claims.”
• “Assess these claims for an EU campaign scheduled to publish on September 1, 2026.”
• “Rewrite the flagged claims conservatively without inventing evidence.”

If the target markets are not provided, the skill evaluates all four available scopes and marks regional applicability as unconfirmed:

• ANY: universal vocabulary and claim-construction risks
• CA: Canadian laws and guidance
• EU: European Union laws and guidance
• UK: United Kingdom laws and guidance

How an audit works

1. The skill determines the input type, intended markets, and assumed publication date.
2. It extracts claim-bearing text while preserving locations such as CSV row and column, workbook sheet and cell, document paragraph, or presentation slide and shape.
3. A bundled Python script performs keyword and pattern triage.
4. The agent reviews each match against the bundled vocabulary, risk taxonomy, detection method, and regional research.
5. It adds contextual risks that keyword matching can miss, including implied claims, imagery, omissions, comparisons, labels, lifecycle scope, and future commitments.
6. It reports findings separately for each applicable jurisdiction and assigns the highest applicable risk as the overall rating.

What the report contains

The resulting audit includes:

• Scope, sources, assumed publication date, and applicable jurisdictions
• Counts by risk level and jurisdiction
• The exact claim and its source location
• Finding code, risk rating, legal or methodological basis, and reasoning
• Practical remediation and a bounded rewrite where appropriate
• Highest-priority fixes
• Limitations and issues requiring legal or technical review

The skill does not invent evidence, percentages, certifications, methods, or environmental benefits. A keyword match is treated as a lead for review, not proof of greenwashing.

Quick test

From the unpacked skill directory, run:

python scripts/analyze_text.py --text "Our eco-friendly product is carbon neutral." --jurisdictions CA EU UK

To audit a file and save Markdown output:

python scripts/analyze_docx.py "document.docx" --jurisdictions CA EU UK --format markdown --out audit.md

Each analyzer supports JSON output by default and Markdown output with --format markdown. See SKILL.md for the full command reference.

Important limitations

• This skill provides compliance triage, not legal advice or a declaration of legal compliance.
• Automated extraction may not capture the full effect of imagery, layout, footnotes, linked evidence, or other visual context.
• Public website analysis does not bypass authentication, paywalls, robots controls, or access restrictions.
• Laws and regulatory guidance can change. Date-sensitive rules should be assessed using the content’s intended publication date.
• Yellow and Red findings should be reviewed with the relevant evidence and escalated to qualified legal counsel when necessary.

See SKILL.md for the agent’s operating instructions and references/ for the governing research, vocabulary, taxonomy, and detection methodology.