LIVE · v1.7 STABLE
BUILD 2026.05.27
ART. 3(1)
AGPL-3.0 · 3 EVAL CASES · 0 ASSERTIONS

IS IT
AN AI SYSTEM
and WHAT
RISK TIER.

Runs the Art. 3(1) definition test, then classifies the risk tier. 7-criteria AI system test, Art. 5 prohibited-practice screening, Annex I + Annex III high-risk routes, Art. 6(3) exception, GPAI systemic-risk check, and Art. 50 transparency triggers — breadth-first across all five tiers.

REG. (EU) 2024/1689   ·   ART. 3(1)   ·   ART. 5   ·   ART. 6   ·   ANNEX I   ·   ANNEX III   ·   ART. 51 GPAI   ·   ART. 53 · 55   ·   ART. 50 TRANSPARENCY   ·   OECD AI FRAMEWORK   ·   ISO 22989   ·   AI OMNIBUS 2026   ·   REG. (EU) 2024/1689   ·   ART. 3(1)   ·   ART. 5   ·   ART. 6   ·   ANNEX I   ·   ANNEX III   ·   ART. 51 GPAI   ·   ART. 53 · 55   ·   ART. 50 TRANSPARENCY   ·   OECD AI FRAMEWORK   ·   ISO 22989   ·   AI OMNIBUS 2026
§ 01 · CAPABILITIES

SIX PRIMITIVES.
One SKILL.

Each capability is documented separately, tested separately, and called by the workflow at the right moment. Compose them or invoke individually.

01
AI SYSTEM TEST
Art. 3(1) definition test across 7 criteria, including the ISO 22989 autonomy scale. Grounded in Commission guidelines and the OECD AI framework.
02
PROHIBITED SCREENING
Silent relevance scoring across all 8 Art. 5 prohibited practices, surfaced as one transparent table. Subject-centric reading per Commission Guidelines.
03
HIGH-RISK ROUTES
Annex I product-safety route (Art. 6(1)) and Annex III application-based route (Art. 6(2)), with the Art. 6(3) narrow-task exception and its profiling re-exception.
04
GPAI PARALLEL TRACK
Independent GPAI model check (Art. 3(63)). Standard GPAI gets Art. 53; systemic-risk GPAI (10^25 FLOP, Art. 51(2)) gets Art. 55. Runs alongside the risk-tier path.
05
SCOPE & OPEN-SOURCE
Art. 2 exclusion analysis (military, personal, R&D, pre-market, ILE) plus dedicated open-source checklists for Art. 2(12) systems and Art. 53(2) GPAI.
06
OMNIBUS DEADLINES
Compliance dates by tier and role, reflecting the AI Omnibus 2026 postponement: Annex III to 2 Dec 2027, Annex I to 2 Aug 2028.
§ 02 · WORKFLOW

FROM INPUT
to ARTEFACT.

Twelve structured steps. The human stays accountable; the skill carries the structure, the citations, and the document trail.

01
Disclaimer
Non-blocking disclaimer noting this is structured guidance, not legal advice, with the Omnibus-postponed effective dates.
02
Prior assessment context
Optionally paste the Assessment Context block from an earlier AI Act skill to pre-fill fields and skip answered questions.
03
Q1 — system description
Capture what the system does, how it works, who uses it, and in what context. This drives every downstream pre-screen.
04
Scope gate (Art. 2)
System-description-informed exclusion check. Surface only plausible exclusions; run the open-source checklist if a free/open-source licence is in play.
05
AI system test (Art. 3(1))
Walk the 7 criteria one at a time, then output the determination table with confidence. If not an AI system, STOP.
06
Prohibited screening (Art. 5)
Score all 8 prohibitions internally; present one table. Deep-dive flagged items. Any hit triggers the prohibited-practice warning and Art. 99(3) penalty.
07
Annex I check
Screen the 18 Annex I product categories. Safety component or product under EU harmonisation law → high-risk under Art. 6(1).
08
Annex III auto-pre-screen
Auto-map the system to the 8 Annex III categories from the description; present a relevance table for user confirmation or override.
09
Art. 6(3) exception
On an Annex III hit, test the 4 narrow-task / complementary conditions. Profiling (Art. 4(4) GDPR) re-triggers high-risk regardless.
10
Route to ai-act-high-risk
On any Annex I or Annex III hit, hand off the depth assessment to the ai-act-high-risk specialist, then return to consolidate the tier.
11
GPAI check (parallel)
Independent of the tier path: is there a GPAI model, and does it cross the 10^25 FLOP systemic-risk threshold? Art. 53 vs. Art. 55.
12
Transparency + dashboard
Apply the Art. 50(1)-(4) triggers, then emit the Classification Report dashboard with tier, legal basis, flags, and the next-skill context block.
§ 03 · ANATOMY

WHAT'S
IN the SKILL.

Single-folder skill. SKILL.md is the runtime spec; references hold the knowledge corpus; evals hold the proof. 

ai-act-classifier/
├── evals  # Test cases + assertions
│   └── evals.json  # 3 cases, 0 assertions
├── references  # Reference corpus
│   ├── ai-system-definition.md
│   ├── art50-transparency.md
│   ├── art6-exception.md
│   ├── case-studies.md
│   ├── compliance-deadlines.md
│   ├── enforcement-framework.md
│   ├── gpai-systemic-risk.md
│   ├── high-risk-annexes.md
│   ├── jurisdiction-requirements.md
│   ├── prohibited-practices.md
│   ├── scope-exclusions.md
│   └── sector-guidance.md
├── CHANGELOG.md  # Version history
├── README.md  # Deployment guide
└── SKILL.md  # Main skill instructions
§ 04 · DEPLOYMENT

INSTALL
and INVOKE.

Two deployment surfaces. The skill auto-triggers on relevant keywords once installed.

CLAUDE.AI USER SKILLS

  1. Settings → Profile → Custom Skills
  2. Upload the entire ai-act-classifier/ folder
  3. Skill auto-triggers on relevant keywords

CLAUDE CODE / MCP

  1. Copy folder to your skills directory:
cp -r ai-act-classifier/ \
   ~/.claude/skills/user/
§ 05 · OUTPUTS

WHAT YOU
get BACK.

Every output is documented, version-pinned, and traceable to its source citation.

AI system determination
Art. 3(1) table scoring all 7 criteria with reasoning, a YES/NO determination, and a stated confidence level.
Prohibited-practice table
All 8 Art. 5 prohibitions with per-item relevance and reasoning; an explicit warning block with Art. 99(3) penalty if any is flagged.
Annex III relevance table
All 8 application categories with relevance assessment and reasoning drawn from the system description.
Classification Report dashboard
Single-page summary: risk tier, classification basis, Art. 6(3) status, scope exclusions, GPAI status, Art. 50 triggers, and flags.
Assessment Context block
Portable hand-off block (system, tier, basis, sector, jurisdiction, GPAI) to paste into ai-act-roles, ai-act-obligations, or ai-act-report.
§ 06 · EVALS

TESTED
before SHIPPED.

Every release runs against a fixed test suite. Assertions check numeric consistency, citation accuracy, and decision-tree branches.

03
Test Cases
0
Assertions
100%
Coverage Required
00
I'm general counsel at a German GmbH (220 employees, IT consulting sector)
0 ASSERTS
01
I'm seeking an AI Act classification for our customer-service AI product
0 ASSERTS
02
I'm legal counsel at a French legal-tech startup
0 ASSERTS
§ 07 · REGULATORY BASIS

WHAT IT
cites.

Every legal verdict resolves to one of these instruments. No invented articles, no synthetic recitals.

EU AI Act (Reg. (EU) 2024/1689)
Art. 2 scope, Art. 3(1) definition, Art. 5 prohibitions, Art. 6 + Annex I/III, Art. 50, Art. 51/53/55.
Commission Guidelines on AI System Definition
Interpretation of the Art. 3(1) 7-criteria test.
Commission Guidelines on Prohibited AI Practices
Subject-centric reading of the Art. 5 prohibitions.
OECD AI Framework
Underlying technical definition behind Art. 3(1).
GPAI Code of Practice
Art. 53 / 55 implementation and the systemic-risk threshold.
Art. 50 Code of Practice
Multi-layered marking framework for transparency obligations.
§ 08 · TRUST

EVERY STEP,
auditable.

The trace is the product. Nothing happens off the record — no hidden tool calls, no silent retrieval, no opaque chain-of-thought.

§
Source-anchored output.
Every criterion verdict, prohibition flag, and tier decision traces to a cited article, Annex entry, or Commission guideline.
VERIFIED
Reproducible decisions.
The Art. 6(4) documentation and Assessment Context block let you recreate and defend a classification later.
IMMUTABLE
†
Test before tiering.
The Art. 3(1) definition test gates everything: no AI system, no risk tier — and a prohibition halts the flow before classification.
ENFORCED
**
EU-native.
Built on Regulation (EU) 2024/1689, Annex III, GPAI thresholds, and the AI Omnibus 2026 timeline. Not retrofitted.
NATIVE
§ 09 · README.MD

EU AI Act System Classifier — Deployment Guide

📄 View the interactive skill page →

See CHANGELOG.md for version history.

Overview

EU AI Act System Classifier — a structured AI Act assessment skill for Claude that provides:

  • Art. 2 scope-exclusion analysis (military, personal use, pure R&D, pre-market, international law enforcement) with system-description-informed targeting
  • Art. 3(1) AI system definition test — 7-criteria analysis grounded in Commission guidelines and the OECD AI framework
  • Open-source exemption checklists — dedicated paths for Art. 2(12) (AI systems) and Art. 53(2) (GPAI models)
  • Art. 5 prohibited-practice screening with subject-centric reading per Commission Guidelines
  • High-risk classification — Annex I product safety route + Annex III application-based route
  • Art. 6(3) exception assessment with Art. 6(4) documentation requirements
  • GPAI model assessment running in parallel with risk-tier classification — standard GPAI (Art. 53) and systemic-risk GPAI (Art. 55)
  • Art. 50 transparency-obligation triggers with Code of Practice multi-layered marking guidance
  • Sector-specific guidance for high-risk Annex III categories (employment, education, biometrics, law enforcement, etc.)
  • Compliance deadline lookup by risk tier and provider/deployer role
  • Classification Dashboard output — single-page summary with legal basis, deadlines, and follow-up actions

File Structure

ai-act-classifier/
├── SKILL.md                              # Main skill instructions (deploy this)
├── CHANGELOG.md                          # Version history
├── evals/
│   └── evals.json                        # 3 test cases (with-skill vs. baseline)
└── references/
    ├── ai-system-definition.md           # Art. 3(1) 7-criteria test + worked examples
    ├── scope-exclusions.md               # Art. 2 exclusion checklists (incl. open-source)
    ├── prohibited-practices.md           # Art. 5 prohibitions — full subject-centric analysis
    ├── high-risk-annexes.md              # Annex I + Annex III routes
    ├── art6-exception.md                 # Art. 6(3) exception with Art. 6(4) docs
    ├── gpai-systemic-risk.md             # Art. 51/53/55 GPAI thresholds and obligations
    ├── art50-transparency.md             # Art. 50 transparency + Code of Practice marking
    ├── sector-guidance.md                # Sector-specific high-risk guidance
    ├── jurisdiction-requirements.md      # Member-state-specific implementation notes
    ├── compliance-deadlines.md           # Deadlines by tier + role
    ├── enforcement-framework.md          # Penalties, market surveillance, AI Office
    └── case-studies.md                   # Worked classification examples

Deployment

Claude.ai (User Skills)

  1. Go to Settings → Profile → Custom Skills (or equivalent)
  2. Upload the entire ai-act-classifier/ folder structure
  3. The skill will auto-trigger when you mention AI Act classification, risk tiers, KI-Verordnung, Annex III, prohibited practices, or GPAI systemic risk

Claude Code / Custom MCP Setup

  1. Copy the ai-act-classifier/ folder to your skills directory: bash cp -r ai-act-classifier/ /path/to/your/skills/user/ai-act-classifier/
  2. Ensure the skill is registered in your configuration

Usage

Quick Start

Describe the AI technology you want to classify:

"We're building a CV-screening tool that ranks job applicants for our HR team. It uses a fine-tuned LLM to score candidates against job descriptions. Do we need to treat this as high-risk under the AI Act?"

The skill will activate and walk you through the classification.

Trigger Phrases

  • "Classify an AI system under the AI Act" / "AI Act risk tier"
  • "Is this an AI system?" / "Art. 3(1)" / "Prohibited practice?"
  • "High-risk classification" / "Annex III" / "Art. 6 exception"
  • "GPAI systemic risk" / "KI-Verordnung" / "Risikoklassifizierung"

For Q&A and article lookup without producing a classification, use the companion ai-act-knowledge skill instead.

Workflow

Phase Description
Phase 1: Scope Gate Art. 2 exclusion analysis (military, personal, R&D, pre-market, ILE) + open-source checklist (Art. 2(12) / Art. 53(2))
Phase 2: AI System Test Art. 3(1) 7-criteria definition test grounded in Commission guidelines and OECD framework
Phase 3: Risk Classification Steps 1–3 in sequence (Art. 5 → Annex I → Annex III + Art. 6(3) exception); Step 4 GPAI assessment runs in parallel; Step 5 Art. 50 transparency check
Phase 4: Classification Dashboard Single-page output: risk tier, legal basis, deadlines, obligations summary, next-step recommendations

Capabilities Summary

Feature Description
Scope Exclusion (Art. 2) System-description-informed targeting — only relevant exclusions surfaced
Open-Source Checklist Dedicated 3-step process for Art. 2(12) AI systems and Art. 53(2) GPAI models
AI System Definition 7-criteria Art. 3(1) test with Commission/OECD guidance
Prohibited Practices Art. 5 screening — subject-centric reading per Commission Guidelines
High-Risk: Annex I Product-safety route (Art. 6(1))
High-Risk: Annex III Application-based route (Art. 6(2)) with Art. 6(3) exception
Art. 6(3) Exception Documented carve-out with Art. 6(4) registration obligations
GPAI Assessment Parallel track — standard GPAI (Art. 53) + systemic-risk GPAI (Art. 55)
Art. 50 Transparency Interaction, synthetic content, emotion recognition, deepfake labeling
Sector Guidance Targeted guidance for Annex III categories (employment, education, biometrics, etc.)
Jurisdiction Notes Member-state-specific implementation requirements
Deadline Lookup Compliance dates by tier (prohibited / high-risk / GPAI / Art. 50 / minimal)
Classification Dashboard Single-page summary output

Regulatory Basis

Document Reference
EU AI Act Regulation (EU) 2024/1689
Commission Guidelines on AI System Definition Art. 3(1) interpretation
Commission Guidelines on Prohibited AI Practices Art. 5 enforcement reading
Commission Guidelines on High-Risk Classification Art. 6 + Annex III
OECD AI Framework Underlying technical definition
GPAI Code of Practice Art. 53/55 implementation
Art. 50 Code of Practice (Transparency) Multi-layered marking framework

License & Disclaimer

This skill provides structured guidance based on the EU AI Act (Regulation (EU) 2024/1689), Commission guidelines, and the OECD AI framework. It does not constitute legal advice. Final classification decisions should involve qualified legal counsel with AI Act expertise.

Licensed under AGPL-3.0 — see LICENSE at the repo root.

Created by Oliver Schmidt-Prietz — OneZero Legal