Nist AI Rmf

NIST AI Risk Management Framework

What this skill does

Applies the NIST AI RMF — by name, by Subcategory, by Action ID — to whatever AI use case, governance question, or assessment the user brings. Three modes; pick one based on the user's question, default to consult if unsure.

1. Consult — fast lookup. "What should I do per the AI RMF for X?" Returns applicable risks (for GenAI) and the relevant Suggested Actions / Subcategories, quoted verbatim. Best for quick gut-check questions.
2. Governance plan — structured plan. "What should our governance plan include per the AI RMF?" Organized around the GOVERN function's Subcategories, with GenAI-specific actions layered in where applicable. Best for standing up or auditing an AI governance program.
3. Assessment — full impact assessment. "Run a NIST AI RMF impact assessment for X." Walks all four functions for one specific system. Best when the user wants a documented artifact.

All three modes share the same source-of-truth: verbatim NIST text in references/. Quote the files; don't invent or paraphrase.

Source and scope

Two NIST publications underlie the skill. The verbatim extracted markdown ships in references/; the raw source HTMLs and maintainer-only re-extraction tooling live outside this distribution.

• NIST AI 100-1 (AI RMF 1.0, January 2023) — the Core framework. Applies to any AI system. Defines Govern, Map, Measure, Manage; their Categories and Subcategories; and seven Trustworthy AI characteristics. Extracted into references/core/.
• NIST AI 600-1 (Generative AI Profile, July 2024) — the GenAI-specific overlay. 12 enumerated GAI risks and 211 Suggested Actions coded GV-X.Y-NNN etc., each mapped to a Core Subcategory. Extracted into references/gai-profile/.

The Core applies to any AI system. The Profile is an overlay on top of the Core for generative systems. So:

• Non-GenAI system → Core only. Don't pull GAI Profile actions; many won't apply.
• GenAI system → Core for the framework + Profile for GenAI-specific risks and actions.
• Mixed pipeline → split per component.

Other NIST AI Profiles exist; they aren't loaded here. If the user asks about one, say so plainly.

Provenance and decline pathways

This skill ships as a standalone skill. The provenance of every claim must be unambiguous to a reader who never saw the conversation.

The skill will:

• Cite verbatim every Subcategory ID and Action ID from references/. The wording in the output must match the file.
• Mark model judgment inline. Applicability calls ("this risk applies here"), operational glosses ("in practice this means…"), role-ownership recommendations, and the final assessment recommendation are model inferences, not NIST statements. Tag the first instance of each kind with [model judgment — verify against system specifics] (or the more specific variants in the templates).
• Distinguish Core vs Profile. Non-GenAI systems never pull GV-/MP-/MS-/MG- action IDs.

The skill will decline to:

• Invent IDs. If a Subcategory or Action ID doesn't appear in references/, it does not exist in NIST's framework. Say so plainly rather than fabricating one.
• Paraphrase NIST text. The framework's authority is the publication. Rewording strips the citation value.
• Assess without input. If the user's system description lacks the detail to assess a Subcategory, list it as an Open item — don't guess.
• Substitute for counsel. NIST is non-binding voluntary guidance. Mandatory regimes (EU AI Act, state AI laws, sector rules) impose actual obligations that may or may not track NIST. Flag the divergence, don't paper over it.

Workflow

In order, every invocation:

1. Read references/README.md first. It's the routing index — it tells you which reference files to load for which question. Don't load files greedily.
2. Gather the question. Identify the AI use case, system, or governance question. If the user's prompt is vague ("what does NIST say about AI?"), ask one clarifying question before drafting — what system, what context, what decision.
3. Decide the mode. Consult / governance plan / assessment. Most queries are consult unless the user explicitly asks for a plan or an assessment.
4. Decide GenAI or not. Foundation models, LLMs, image/audio/video/text generators, RAG over a generative core — GenAI. Classifiers, regressors, recommenders, anomaly detectors, traditional ML — not GenAI (use Core only).
5. Load only the reference files the question needs, per references/README.md.
6. Load the relevant output template from references/templates/<mode>.md when drafting output.
7. Produce output per the template with verbatim citations and the provenance markers described above.

Mode 1 — Consult

When to use: the user is asking "what should we do?" or "what does NIST say about?" with a specific system or scenario in mind. Fast turnaround. Not a deliverable artifact.

Procedure:

1. From the system description, identify:
   • System type (GenAI? non-GenAI? mixed?).
   • For GenAI: which of the 12 GAI risks plausibly apply. Use gai-profile/risks.md + crosswalk.md. Be honest — if a risk obviously doesn't apply (e.g., CBRN for a customer-service chatbot), say so and exclude it. Don't pad.
   • For any system: which of the Core Subcategories most directly apply. Usually 4–10, not all of them.
2. Pull the relevant Suggested Actions (GenAI) or Subcategory statements (non-GenAI) into a table. Group by function.
3. Surface 2–4 follow-up questions the user can't answer from the framework alone (e.g., "What's your incident response capacity?"). The framework points; the user fills in.

Output template: references/templates/consult.md — load when drafting.