ai-transparency-writer

AI Transparency Writer

Drafts compliant, publication-ready transparency notices, model cards, and system disclosures. The output is actual document text — not a checklist — scoped to the framework and role that applies to the user's situation.

---

Step 1 — Gather Context

Collect the following before drafting. Infer what you can from the conversation; ask only for what's genuinely missing:

1. What does the AI system do? (plain-language description of function and outputs)
2. Who interacts with or is affected by it? (end users, operators, impacted individuals)
3. What is the user's role? — provider (develops/places on market) or deployer (uses in own context)?
4. Which framework to target?
   • EU AI Act Article 13 (transparency for high-risk AI)
   • NIST AI RMF model card / system disclosure
   • Both
5. Has the system been classified under the EU AI Act? If not and the user is targeting Article 13, recommend running eu-ai-act-classifier first — Article 13 only applies to high-risk systems, and the obligations differ by role.

If the framework is ambiguous, ask: "Should I draft an EU AI Act Article 13 notice, a NIST-aligned model card, or both?"

---

Step 2 — Select Mode and Load References

Target framework	Load
EU AI Act Article 13	references/article-13-requirements.md
NIST AI RMF model card	references/model-card-template.md
Both	Both files
Ambiguous	Ask (see Step 1)

Also load references/disclosure-examples.md for language patterns and worked examples regardless of mode — it helps produce natural, accurate prose rather than boilerplate.

---

Step 3 — Draft the Document

Generate a complete, publication-ready draft using the structure and requirements from the loaded reference files. The output should be real text the user can drop into a document, website, or system interface — not a template with blanks.

EU AI Act Article 13 — Transparency Notice

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📄 AI SYSTEM TRANSPARENCY NOTICE
Prepared under: EU AI Act Article 13
System: [system name]
Deployer / Provider: [org name]
Date: [today]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⚠️ ADVISORY NOTICE
───────────────────
This document is drafted by an AI skill and is provided for informational and
governance support purposes only. It does not constitute legal advice or a formal
compliance determination. Do not publish or rely on this notice as a substitute for
review by qualified legal counsel or a licensed compliance professional with
jurisdiction-specific expertise.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

[Draft notice text — see references/article-13-requirements.md for required fields
and references/disclosure-examples.md for example language. Write full prose
paragraphs for each mandatory section. Do not leave placeholder brackets in the
final output — use [ field — to be confirmed ] only where the user genuinely
hasn't provided the information.]

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

NIST AI RMF — Model Card / System Disclosure

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🗂️  AI SYSTEM MODEL CARD
Framework: NIST AI RMF 1.0 (GOVERN function)
System: [system name]
Organization: [org name]
Version: [version / date]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⚠️ ADVISORY NOTICE
───────────────────
This document is drafted by an AI skill and is provided for informational and
governance support purposes only. It does not constitute legal advice or a formal
compliance determination. Do not publish or rely on this notice as a substitute for
review by qualified legal counsel or a licensed compliance professional with
jurisdiction-specific expertise.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

[Draft model card text — see references/model-card-template.md for section
structure and references/disclosure-examples.md for example language. Write
full prose for each section; do not leave sections blank.]

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Drafting principles:

• Write in plain language accessible to affected individuals — not legal jargon
• Be specific: generic boilerplate fails compliance reviews and erodes user trust
• Where information is unknown, flag explicitly (e.g., "[ human oversight contact — to be confirmed ]") rather than omitting or writing vague filler
• For deployers under EU AI Act: focus on use-context disclosures, not technical model details (that's the provider's job)
• For providers drafting model cards: focus on intended uses, limitations, and risk mitigations

---

Step 4 — Compliance Self-Check

After drafting, run a quick internal check:

For Article 13 notices:

• Verify all mandatory fields from references/article-13-requirements.md are present
• Flag any field where user input was insufficient and a placeholder was used
• Note whether the system's risk tier has been confirmed

For NIST model cards:

• Verify all sections from references/model-card-template.md are addressed
• Flag any section left thin due to missing information

Present the draft followed by a brief compliance note:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
✅ COMPLIANCE NOTE
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Mandatory fields covered: [list]
Fields needing confirmation: [list any placeholders or gaps]
Risk tier confirmed: [Yes / No — recommend eu-ai-act-classifier if No]
Recommended next step: [e.g., legal review, hitl-compliance-gate sign-off]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

---

Step 5 — Offer Outputs and Next Steps

After presenting the draft and compliance note, offer:

1. Save as Markdown — write the notice to a .md file for version control or sharing
2. Export as Word document — hand off to the docx skill to produce a formatted .docx
3. Compliance gate — trigger hitl-compliance-gate with EU AI Act pre-selected before publishing or distributing to affected users
4. Refine a section — iterate on specific sections inline if the user wants improvements
5. Draft both frameworks — if only one was produced, offer to generate the other

---

Key Principles

• Advisory only — All outputs produced by this skill are for informational and governance support purposes only. They do not constitute legal advice, regulatory advice, or a formal compliance determination. No output from this skill should be relied upon as a substitute for advice from qualified legal counsel, a licensed compliance professional, or a certified auditor with jurisdiction-specific expertise.
• Documents, not checklists — the output must be readable and publishable prose, not a list of boxes to fill in later
• Plain language is a compliance requirement — Article 13 notices must be understandable by affected individuals; dense legalese is itself a compliance failure
• Specificity over safety blankets — "the AI may make errors" tells users nothing; "the AI's credit scoring output may underweight recent income changes and should be reviewed by an officer before a final lending decision" is actionable
• Role determines scope — deployers disclose use-context, purpose, and human oversight contacts; providers disclose model characteristics, training data categories, and limitations